[27] Asymmetrical body structures producing symmetrical running in quadrupedal mammals

Mau Adachi¹, Tomoya Kamimura¹, Yuichi Ambe², Yasuo Higurashi³, Naomi Wada³, Fumitoshi Matsuno⁴, and Shinya Aoi¹

[1] Department of Mechanical Science and Bioengineering, Graduate School of Engineering and Science, The University of Osaka, Osaka, Japan
[2] Department of Electrical, Systems, and Control Engineering, Graduate School of Advanced Science and Engineering, Hiroshima University, Hiroshima, Japan
[3] Laboratory of System Physiology, Joint Faculty of Veterinary Medicine, Yamaguchi University, Yamaguchi, Japan
[4] Department of Electronics and Information Systems Engineering, Osaka Institute of Technology, Osaka, Japan

Quadrupedal mammals often exhibit highly symmetrical locomotion patterns during running, despite their anatomically asymmetric bodies. In this study, we developed a simplified dynamical model incorporating multiple foreaft asymmetries (center-of-mass offset and leg stiffness) and analyzed their effects on quadrupedal running dynamics in the sagittal plane. Our results demonstrate that when specific relationships between asymmetry parameters are satisfied, the model yields symmetrical running patterns with enhanced gait performance, such as reduced vertical impulses. These findings suggest that quadrupedal mammals may coordinate internal asymmetries to achieve high-performance locomotion.

[42] Static Pose Optimization to Reduce Connector and Joint Loads in a Modular-Limbed Quadruped with Position-Controlled Joints

Yuichi Ambe¹, Takashi Takuma², Tomohiro Hayakawa³, and Fumitoshi Matsuno²

[1] Graduate School of Advanced Science and Engineering, Hiroshima University, Higashi-hiroshima, Japan
[2] Department of Electronics and Information Systems Engineering, Osaka Institute of Technology, Osaka, Japan
[3] Department of Mechanical Engineering, Shizuoka University, Hamamatsu, Japan

We are developing a modular robot for lunar exploration that can reconfigure its modules through detachable connectors. In its quadruped configuration, minimizing loads on connectors and joints is important for protecting the hardware and reducing energy consumption. However, this is challenging because the limb joints use position-controlled servos instead of torque-controlled actuators. This study proposes a method to compute connector and joint torques using potential energy minimization under foot position constraints. We also demonstrate a static posture optimization that reduces these loads when the quadruped swings one leg while the other three legs remain grounded.

[26] Development of a Remote Monitoring User Interface for Self-organized Modular Robots for Lunar Exploration

Haruki Aoyama¹, Kotaro Kanazawa¹, Ching Wen Chin², Guang Yang³, Xixum Wang³, Ryohei Michikawa⁴, Noritaka Sato¹, Motoyasu Tanaka², Akio Noda⁵, and Fumitoshi Matsuno³

[1] Graduate School of Engineering, Nagoya Institute of Technology, Aichi, Japan
[2] Graduate School of Engineering, The University of Electro-Communications, Tokyo, Japan
[3] Graduate School & Faculty of Engineering, Osaka Institute of Technology, Osaka, Japan
[4] Graduate School of Engineering, Kyoto University, Kyoto, Japan
[5] Graduate School & Faculty of Robotics & Design Engineering, Osaka Institute of Technology, Osaka, Japan

This paper presents the development and implementation of a remote-monitoring user interface for Selforganized modular robots designed for lunar exploration and base construction. The system visualizes the dynamic connection states and roles of heterogeneous modules—such as body, limb, wheel, and gripper units—using continuously updated sensor data and ROS-based communication protocols. A hierarchical leader–follower relationship among modules is dynamically managed and displayed through the interface. Experiments demonstrate the interface’s ability to depict autonomous module reconfiguration, grasping actions, and navigation over lunar-like terrain. The proposed interface enables users to intuitively understand the robot’s configuration and operation status, contributing to autonomous, resilient robotic missions in remote and harsh environments such as the Moon.

[68] Recording and Analysis of Interactive Human Reaching Motions with Cooperative Collision Avoidance

Simon Blaue¹, Minija Tamosiunaite^1,2, and Florentin Wörgötter¹

[1] Department of Physics, III. Institute, Georg-August University, Göttingen, Germany
[2] Vytautas Magnus University, Kaunas, Lithuania

This study investigates interactive human reaching motions with a focus on cooperative collision avoidance strategies. We present a dataset of reaching tasks recorded using a multi-camera motion capture system synchronized with eye-tracking devices. Participants performed coordinated reaching movements across a shared workspace, whose trajectories we analyzed for deviations from their individual baselines. Key kinematic metrics, including movement du- ration, curvature, tilt, velocity profiles, and leader-follower dynamics, were quantitatively evaluated. Results indicate that interactive motions exhibit prolonged durations (up to 400 ms), increased curvature (fourfold average increase), and context-dependent tilt adjustments compared to solo movements. Notably, systematic leader-follower strategies were ab- sent in 82 % of trials, suggesting reliance on real-time reactive adjustments rather than preemptive waiting. Dynamic Movement Primitives (DMPs) were employed to encode trajectories, revealing distinct motion characteristics through principal component analysis (PCA). These findings provide insights into human cooperative motion planning, with im- plications for robotics and human-robot interaction frameworks requiring adaptive collision avoidance

[58] Aerial Interception of Multiple Drones by Multiple Distributed Chasers with Batch Deployments

N. Charuvajana¹, X.Y. Tan¹, P. Rithburi², S. Srigrarom³, B.C. Khoo³, and F. Holzapfel⁴

[1] National University of Singapore, Computer Science Dept, Singapore
[2] Technical University of Munich Asia, Singapore
[3] National University of Singapore, Mechanical Engineering Dept, Singapore
[4] Technical Universtiy of Munich, Institute of Flight Systems Dynamics, Germany

In this paper, we investigate the problem of air-to-air drone interception for counter-UAS operations, using strategic surrounding approaches and sequences of deployments. The intruding drones may come in large numbers and are modeled by Boid Algorithm, analogous to flocks of birds. They assumed to destine towards the goal, of which the interceptors need to protect. The fleet of interceptors (agents) are assumed to be limited numbers and are equal to or smaller than the intruding drones, and are spaced out in batches for cover large volume either by 1) zonal (distance based) deployment or 2) scheduled (fixed time) deployments. During the interception, we employed a greedy strategy as a taskallocation method to break up, herd, and encircle the intruders. This framework is modeled as matching and optimization problem. The preliminary mix-integer, non-linear problem (MINLP) formulations are based on probability of interception and resource readiness. The interceptors are designed to be deployed in sequences of batches (groups) such that they can cover larger area, and allow follow-up action, should the first batch is not sufficient. Preliminary results show that the batches deployments gives higher success rate than the single deployment. The zonal deployment performs better for moderate number of intruders and interceptors (10-30), whereas the scheduled (fixed time) deployment performs better for larger number of intruders and interceptors (≥ 30).

[69] A Differential Steering Mechanism for Directional Adaptation and Self-Organized Locomotion under Decentralized Control in Legged Robots

Thirawat Chuthong¹, and Poramate Manoonpong¹

[1] Bio-inspired Robotics and Neural Engineering Lab, School of Information Science and Technology, Vidyasirimedhi Institute of Science and Technology, Wangchan Valley 555 Moo 1, Payupnai, Wangchan, Rayong, Thailand

Legged robots offer remarkable adaptability in navigating complex environments, but achieving reliable direction control remains a challenge, especially under decentralized controls that avoid high computation, prior training or detailed modeling. While such controls enable rapid gait formation using only minimal proprioceptive feedback and robot-environment interaction, they often result in unstable heading behavior, such as unintended turns or a random walking direction. In this study, we propose a differential adaptive steering control mechanism that operates in conjunction with decentralized locomotion control to stabilize and guide the robot’s heading using only body orientation in the yaw angle. Our method modulates the leg movement amplitude to achieve target-aligned directional adaptation. We validate our approach on a simulated stick insect-like robot on both flat and rough terrains of varying roughness. The results demonstrate robust and directionally stable locomotion, enabling the robot to follow straight and complex paths.

[15] Autonomous Role Allocation in Multi-Agent Systems: Inspired by Ant Division of Labor

Makoto Eguchi¹, and Kazuaki Yamada²

[1] Graduate School of Science and Engineering, Toyo University, Saitama, Japan
[2] Department of Mechanical Engineering, Toyo University, Saitama, Japan

This paper proposes a novel autonomous role allocation method for multi-agent systems, inspired by the division of labor mechanisms observed in ants. Multi-agent systems have no central control system. Instead, many autonomous agents form an ordered system through interactions with their environments and neighbors. This inherent structure is expected to yield high robustness, flexibility, fault tolerance, and scalability. Inspired by the division of labor mechanisms observed in social insects, such as ants, we propose an autonomous role allocation method. Specifically, it introduces three key components: the response threshold model, the response threshold variance model, and the response thresholds specialization model. These models are directly derived from the known the division of labor behaviors in ant colonies. Through simulation experiments, we demonstrate the proposed method’s application to task assignment problems. Our results show its ability to adaptively allocate roles in response to internal and external system changes, facilitated by local communication among agents.

[22] Quantifying swarm reorganization in soldier crabs under external threats

Claudio Feliciani¹, Zeynep Yacel², Hisashi Murakami³, Yuto Uesugi⁴, Takenori Tomaru³, Tamao Maeda⁵, and Sakurako Tanida¹

[1] Graduate School of Engineering, The University of Tokyo, Tokyo, Japan
[2] Department of Environmental Sciences, Informatics and Statistics, Ca’ Foscari University of Venice, Venice, Italy
[3] Faculty of Information and Human Science, Kyoto Institute of Technology, Kyoto, Japan
[4] Graduate School of Science, University of Tokyo, Tokyo, Japan
[5] Research Center for Integrative Evolutionary Science, Graduate University of Advanced Studies, Hayama, Japan

This study investigates the collective behavior of soldier crab swarms in response to external threats. Using aerial video observations and custom computer vision techniques, we identify how swarm density influences behavioral shifts. Preliminary analysis reveals distinct responses based on local organization: isolated or loosely grouped crabs adaptively reorganize, while dense swarms remain largely unresponsive. A framework for detecting swarm boundaries is introduced, enabling future quantitative studies relevant to both biological swarms and swarm robotics.

[05] Development of Wheels and a Tail for a Two-Wheeled Lunar Rover

Yuto Fukao¹, Yuya Shimizu², Ryohei Michikawa¹, Kiona Hosotani¹, Ryusuke Fujisawa³, Tetsushi Kamegawa², and and Fumitoshi Matsuno⁴

[1] Faculty of Engineering, Kyoto University, Kyoto, Japan
[2] Graduate School of Environment, Life, and Natural Science and Technology, Okayama University, Okayama, Japan
[3] Graduate School and Faculty of Environmental Engineering, The University of Kitakyushu, Fukuoka, Japan
[4] Graduate School and Faculty of Engineering, Osaka Institute of Technology, Osaka, Japan

In this study, the design and development of the wheels and tail for a two-wheeled rover intended for lunar exploration are presented, together with the results of driving experiments conducted in a sandy environment that simulates the lunar surface. The wheels incorporate circumferential treads that push against the sand to generate propulsion; four wheel variants with different tread widths and heights were developed, and driving tests were performed on both level ground and inclined terrain, after which the results obtained on the two terrains were compared and analyzed. The tail, which is pressed into the sand during driving to prevent rotation of the main body, was realized as an omni-ball configuration capable of omnidirectional rotation so as to accommodate both translational and rotational motions of the rover, and its effectiveness was verified through driving experiments on sand.

[08] Graph Grammar Definition Language GGDL and its application

Koki Harada^1,2, Ryo Ariizumi², Xixun Wang³, Ryota Kinjo³, and Fumitoshi Matsuno³

1 Nagoya University, Nagoya, Japan (E-mail: harada.koki.s8@s.mail.nagoya-u.ac.jp) 2 Tokyo University of Agriculture and Technology, Tokyo, Japan 3 Osaka Institute of Technology, Osaka, Japan

A graph grammar is a set of graph rewriting rules for graph generation. Graph grammars have a wide scope of application including the robot science. However, there are few concrete formats to describe a graph grammar, which prevents researchers sharing, revising and reusing a grammar. In this paper, we present a novel computer language GGDL for description of a graph grammar, and show an application to the auto-generation of robot structures as an example.

[64] Stability change in a quadruped robot by stride frequency via neuromechanical resonance

Takatoshi Hashimoto¹, Tomoya Kamimura¹, Sota Tsuchida¹, Tomoki Nakagawa¹, Yuichi Ambe², Mau Adachi¹, and Shinya Aoi¹

[1] Department of Mechanical Science, Graduate School of Engineering Science, The University of Osaka, Osaka, Japan
[2] Graduate School of Advanced Science and Engineering, Hiroshima University, Hiroshima, Japan

The study presents a quadruped robot with four pantograph-shaped legs, each actuated at both ends and controlled to mimic linear and torsional springs. Locomotion is generated using a two-layer central pattern generator (CPG) system: the rhythm generator (RG) layer produces rhythmic signals, and the pattern formation (PF) layer converts them into leg spring commands based on kinematics. By modulating phase oscillators and using sensory feedback, the robot can produce walking, trotting, and pacing gaits from a single phase difference between ipsilateral oscillators.

[24] Task allocation strategy for reconfigurable modular robots based on trophallaxis and cannibalism

Tomohiro Hayakawa¹, Ryusuke Fujisawa², Yuki Tanigaki³, and Fumitoshi Matsuno³

[1] Shizuoka University, Hamamatsu, Japan
[2] University of Kitakyushu, Kitakyushu, Japan
[3] Osaka Institute of Technology, Osaka, Japan

In this study, we consider a task allocation problem for reconfigurable modular robots as a swarm. The performance of reconfigurable modular robots depends on the number of modules in a cluster; thus, the number of modules in each cluster should be adaptively adjusted. In this study, we incorporate the behavior of trophallaxis and cannibalism in natural creatures into the task allocation for the swarm of reconfigurable modular robots. Through multiagent simulation, we found that both trophallaxis and cannibalism behavior enables rapid task execution for the swarm of reconfigurable modular robots.

[48] Effects of Tail Stiffness on Swimming Performance in Individual and Schooling Conditions

Hun Jang^1,2,3, and Liang Li^1,2,3,4

[1] Department of Collective Behaviour, Max Planck Institute of Animal Behavior, Konstanz, Germany
[2] Centre for the Advanced Study of Collective Behaviour, University of Konstanz, Konstanz, Germany
[3] Department of Biology, University of Konstanz, Konstanz, Germany
[4] Department of Computer and Information Science, University of Konstanz, Germany

Fish use efficient swimming patterns both individually and in schools, benefiting from vortices generated by themselves and neighboring fish, with factors like tail motion, positioning, and phase differences influencing performance. This study explores the underexamined role of passive joints in enhancing hydrodynamic interactions and energy efficiency, aiming to inform both biological understanding and bio-inspired robotic fish design.

[20] Position-Based Flocking for Robust Alignment

Hossein B. Jond¹

[1] Department of Cybernetics, Czech Technical University in Prague, Prague, Czechia

This paper presents a position-based flocking model for interacting agents, balancing cohesion-separation and alignment to achieve stable collective motion. The model modifies a velocity-based approach by approximating velocity differences using initial and current positions, introducing a threshold weight to ensure sustained alignment. Simulations with 50 agents in 2D demonstrate that the position-based model produces stronger alignment and more rigid and compact formations compared to the velocity-based model. The alignment metric and separation distances highlight the efficacy of the proposed model in achieving robust flocking behavior. The model’s use of positions ensures robust alignment, with applications in robotics and collective dynamics.

[28] Dynamical Effects of Asymmetry Body Flexibility on Cheetah Running Gait

Tomoya Kamimura¹, Yuya Oshita², Mau Adachi¹, Yuichi Ambe³, Akihito Sano², Naomi Wada⁴, Shinya Aoi¹, and Fumitoshi Matsuno⁵

[1] Department of Mechanical Science and Bioengineering, Graduate School of Engineering and Science, The University of Osaka, Osaka, Japan
[2] Department of Electrical and Mechanical Engineering, Nagoya Institute of Technology, Aichi, Japan
[3] Department of Electrical, Systems, and Control Engineering, Graduate School of Advanced Science and Engineering, Hiroshima University, Hiroshima, Japan
[4] Laboratory of System Physiology, Joint Faculty of Veterinary Medicine, Yamaguchi University, Yamaguchi, Japan
[5] Department of Electronics and Information Systems Engineering, Osaka Institute of Technology, Osaka, Japan

Cheetah’s high-speed running involves remarkable spine bending. In this study, we focused on the dynamical effect of asymmetric body flexibility on running. We proposed a simple dynamical model with asymmetry body flexibility and searched its periodic solutions numerically. The obtained solutions showed the mechanism under which the asymmetry of flexibility reduces the ground reaction force on the legs and explains the superiority of the cheetah’s running gait.

[60] Ant-inspired Walling Strategies for Scalable Swarm Separation: Reinforcement Learning Approaches Based on Finite State Machines

Shenbagaraj Kannapiran¹, Elena Oikonomou², Albert Chu², Spring Berman¹, and and Theodore P. Pavlic^2,3

[1] School for Engineering of Matter, Transport and Energy, Arizona State University, Tempe, AZ 85287, USA
[2] School of Computing and Augmented Intelligence, Arizona State University, Tempe, AZ 85281, USA
[3] School of Life Sciences, Arizona State University, Tempe, AZ 85281, USA

In natural systems, emergent structures often arise to balance competing demands. Army ants, for example, form temporary “walls” that prevent interference between foraging trails. Inspired by this behavior, we developed two decentralized controllers for heterogeneous robotic swarms to maintain spatial separation while executing concurrent tasks. The first is a finite-state machine (FSM)-based controller that uses encounter-triggered transitions to create rigid, stable walls. The second integrates FSM states with a Deep Q-Network (DQN), dynamically optimizing separation through emergent “demilitarized zones.” In simulation, both controllers reduce mixing between subgroups, with the DQN-enhanced controller improving adaptability and reducing mixing by 40–50% while achieving faster convergence.

[54] Probabilistic Path Selection Method in Planetary Exploration: Improving Terrain Adaptability through Tethered Cooperative Rover Systems

Clive Jancen Kawaoto¹, Tenta Suzuki¹, Mao Tobisawa¹, Yuki Itoh¹, Kaito Kumagae¹, Johei Matsuoka¹, and and Kiyohiko Hattori²

[1] Tokyo University of Technology, 1404-1, Katakurama-chi, Hachioji-shi, Tokyo, Japan
[2] Tokyo Denki University, 5 Senju Asahi-cho, Adachi-ku, Tokyo, Japan

In planetary exploration, exploration rovers must perform autonomous path selection in unknown terrain environments. However, there is a challenge where exploration efficiency decreases due to conservative path selection that detours around even traversable hazardous areas to avoid the risk of getting stuck. To address this challenge, this study considers an autonomous cooperative navigation system using two small two-wheeled rovers connected by a wire. It proposes a probabilistic exploration strategy to achieve exploration independent of terrain patterns. This method increases learning opportunities for terrain by attempting to enter groove terrain classified as hazardous with a certain probability, enabling adaptive path learning. Previous research suffered from learning results being significantly influenced by the order of groove appearance, particularly in environments with many deep grooves during the early exploration phase, where even traversable shallow grooves were avoided. To evaluate the effectiveness of the proposed method, we established four different arrangement patterns and conducted evaluations through simulations with nine combinations of start and goal positions. As a result, we achieved improvements of 6.4% in travel distance improvement rate and 3.4% in path efficiency, particularly in environments where many deep grooves were arranged during the early exploration phase. This demonstrates the feasibility of an autonomous exploration system that is less dependent on groove arrangement patterns.

[30] Building a Motion Embedding Space for Annotation-Free Animal Behavior Recognition

Mayu Kikuchi¹, Yasumasa Tamura², and Masahito Yamamoto²

[1] Graduate School of Information Science and Technology, Hokkaido University, Japan
[2] Faculty of Information Science and Technology, Hokkaido University, Japan

Animal behavior recognition is essential for ensuring animal welfare and effective management in zoos. However, conventional approaches require large-scale labeled data for all target behaviors, limiting their scalability to new environments and behaviors. To address this challenge, we propose a dual-stream architecture with Gated Fusion that adaptively integrates visual cues from RGB videos with motion information from optical flow. We trained our model on the large-scale Animal Kingdom dataset and evaluated it on an out-of-domain dataset consisting of surveillance footage of polar bears provided by Sapporo Maruyama Zoo. Experimental results show that our Gated Fusion model achieved the highest performance on key clustering metrics such as the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI), substantially outperforming models that rely on a single modality (either RGB or optical flow). These findings highlight the synergistic benefit of combining appearance and motion, and demonstrate the potential of our approach for robust, unsupervised behavior recognition in unseen environments.

[07] VariAntNet: Learning Decentralized Control of Multi-Agent Systems

Yigal Koifman¹, Erez Koifman¹, Eran Iceland¹, Ariel Barel¹, and and Alfred M. Bruckstein¹

[1] Department of Computer Science, Technion †“ Israel Institute of Technology, Haifa, Israel

A simple multi-agent system can be effectively utilized in disaster response applications, such as firefighting. Such a swarm is required to operate in complex environments with limited local sensing and no reliable inter-agent communication or centralized control. These simple robotic agents, also known as Ant Robots, are defined as anonymous agents that possess limited sensing capabilities, lack a shared coordinate system, and do not communicate explicitly with one another. A key challenge for simple swarms lies in maintaining cohesion and avoiding fragmentation despite limited-range sensing. Recent advances in machine learning offer effective solutions to some of the classical decentralized control challenges. We propose VariAntNet, a deep learning-based decentralized control model designed to facilitate agent swarming and collaborative task execution. VariAntNet includes a preprocessing stage that extracts geometric features from unordered, variable-sized local observations. It incorporates a neural network architecture trained with a novel, differentiable, multi-objective, mathematically justified loss function that promotes swarm cohesiveness by utilizing the properties of the visibility graph Laplacian matrix. VariAntNet is demonstrated on the fundamental multi-agent gathering task, where agents with bearing-only and limited-range sensing must gather at some location. VariAntNet significantly outperforms an existing analytical solution, achieving more than double the convergence rate while maintaining high swarm connectivity across varying swarm sizes. While the analytical solution guarantees cohesion, it is often too slow in practice and fails to meet the required convergence time. In time-critical scenarios, such as emergency response operations where lives are at risk, rapid convergence is crucial, making slower analytical methods impractical and justifying the loss of some agents within the swarm. This paper presents and analyzes this trade-off in detail.

[17] Modeling and Analysis of Locust Swarms with OR35-Deficient Individuals

Haruto Maeda¹, and Masaki Ogura¹

[1] Graduate School of Advanced Science and Engineering, Hiroshima University, Hiroshima, Japan

Locust plagues have posed a severe and persistent threat to agriculture for thousands of years, yet no definitive or sustainable solution has been found. While insecticides are currently the primary control method, effective intervention is limited to the nymph stage, and concerns remain regarding their impact on ecosystems and human health. These limitations highlight the need for alternative strategies for locust swarm management. Previous studies have identified 4-vinylanisole (4VA), a volatile compound emitted by the migratory locust (Locusta migratoria), as an aggregation pheromone, and have shown that the olfactory receptor OR35 specifically detects 4VA. They have also reported that individuals lacking OR35 show a reduced attraction to 4VA. Focusing on this aggregation pheromone and its receptor OR35, this study analyzes the effects of OR35-deficient individuals on swarm dynamics through simulation. To this end, we developed a mathematical model designed to capture the influence of OR35-deficient individuals on swarm dynamics, building upon established self-propelled particle models and empirical findings on locust movement patterns. Our simulation results show that, within the framework of this model, varying the proportion of OR35-deficient individuals leads to distinct behavioral patterns of swarm dynamics, suggesting a possible direction for future exploration of non-insecticidal control strategies.

[19] Self-Reconfigurable Modular Robots for Lunar Exploration and Base Construction †” Overview and Interim Results of a Research Theme within the Moonshot Project

Fumitoshi Matsuno¹, Xixun Wang¹, Guang Yang¹, Kiona Hosotani², Ryohei Michikawa², Tetsushi Kamegawa³, Yuya Shimizu³, Motoyasu Tanaka⁴, Ching Wen Chin⁴, Ryusuke Fujisawa⁵, Ryo Ariizumi⁶, Koki Harada⁷, Tomohiro Hayakawa⁸, Yuichi Ambe⁹, Akio Noda¹, Hiroshi Oku¹, Takashi Takuma¹, Ryota Kinjo¹, Yuki Tanigaki¹, Noritaka Sato¹⁰, and and Kotaro Kanazawa¹⁰

[1] Osaka Institute of Technology, Osaka, Japan
[2] Kyoto University, Kyoto, Japan
[3] Okayama University, Okayama, Japan
[4] The University of Electro-communications, Chofu, Japan
[5] The University of Kitakyushu, Kitakyushu, Japan
[6] Tokyo University of Agriculture and Technology, Koganei, Japan
[7] Nagoya University, Nagoya, Japan
[8] Shizuoka University, Hamamatsu, Japan
[9] Hiroshima University, Higashihiroshima, Japan
[10] Nagoya Institute of Technology, Nagoya, Japan

We have developed self-reconfigurable modular robots designed for constructing base stations and conducting exploratory missions on the Moon. This paper outlines the objectives of our project and reports on the current progress. We describe the design and development of key components, including the module connector, body module, limb module, end-effector module, and user interface. Furthermore, we present the controller design and the task allocation strategy employed for swarm-based modular robotic systems.

[63] Diffusion-model-based Coordination of Multi-agent Systems: Achieving Formations Similar to–but Distinct from–References

Aki Matsutaka¹, Shun-ichi Azuma¹, and Ikumi Banno¹

[1] Graduate School of Informatics, Kyoto University, Kyoto, Japan

This study focuses on decentralized multi-agent systems, aiming to create formations that are similar to, but not exactly the same as, predefined reference formations. The proposed distributed control framework uses a diffusion model within each controller to guide agents toward their desired positions. Simulation results with 1,000 agents forming variations of the MNIST digit “2” demonstrate the method’s potential for automatic formation generation.

[25] Development of Robot Module Connectors with Dustproof and Passive Error Compensation Functions Using Metal Sheets

Ryohei Michikawa¹, Kiona Hosotani¹, Xixun Wang², Yuki Takagi², Haruho Mitsunaga³, Mihiro Nakabayashi⁴, Guang Yang², Ryusuke Fujisawa⁵, Akio Noda³, Tomohiro Hayakawa⁶, Hiroshi Oku², and Fumitoshi Matsuno²

[1] Department of Mechanical Engineering and Science, Kyoto University, Kyoto, Japan
[2] Department of Electronics and Information Systems Engineering, Osaka Institute of Technology, Osaka, Japan
[3] Department of Robotics and Design, Osaka Institute of Technology, Osaka, Japan
[4] Faculty of Engineering, Toyama University, Toyama, Japan
[5] Faculty of Environmental Engineering, Kitakyushu University, Kitakyushu, Japan
[6] Department of Mechanical Engineering, Shizuoka University, Hamamatsu, Japan

In extreme environments such as the Moon, modular robot connectors must resist dust infiltration and tolerate significant alignment errors. We developed a dustproof connector that fully isolates the drive mechanism using folded metal sheets. Additionally, we developed a mechanism that absorbs posture errors during coupling using passive joints with controllable ranges of motion. The proposed mechanism was mounted on an actual robot, and its effectiveness was verified.

[57] A Neuro-musculoskeletal Model-based Algorithm for Virtual Weight Presentation via Electrical Muscle Stimulation

Ryohei Michikawa¹, Hiroshi Yokoi², and Fumitoshi Matsuno³

[1] Department of Mechanical Engineering and Science, Kyoto University, Kyoto, Japan
[2] Department of Mechanical and Intelligent Systems Engineering, The University of Electro-Communications, Tokyo, Japan
[3] Department of Electronics and Information Systems Engineering, Osaka Institute of Technology, Osaka, Japan

Electrical stimulation of antagonist muscles with surface electrodes can evoke the sensation of weight, yet the mapping between stimulus intensity and perceived weight—especially across postures involving multiple muscles— remains unclear. We hypothesize that central signals are the primary source of weight perception and model this percept when several muscles are stimulated simultaneously. Building on this model, we propose an algorithm that, for an arbitrary static posture, allocates stimulus intensities so that the resulting distribution of central signals replicates that produced when an actual object is supported. Simulations confirm that the algorithm reproduces the central-signal distribution observed during physical loading.

[03]

Konstantin Möller

[59] Isomorphism Determination Method for Modular Robots Using Graph Neural Networks

Ryuusei Nishii¹, Kenichiro Satonaka¹, Ryota Kinjo¹, Hiroshi Oku¹, Yuki Tanigaki¹, Tomohiro Shimomura¹, Guang Yang¹, Xixun Wang¹, Fumitoshi Matsuno¹, and Ryo Ariizumi²

[1] Osaka Institute of Technology, Osaka, Japan
[2] Tokyo University of Agriculture and Technology, Tokyo, Japan

In this study, we explore the propose of an AI system that leverages GNN to efficiently search for transformation paths for modular robots. Determining isomorphism between robot configurations is essential for enabling efficient searching of reconfiguration planning. We first train a GNN-based model to recognize whether two robot states are isomorphic. Next, we use this trained isomorphism determination GNN as the foundation for transfer learning, aiming to build an AI system that can judge whether one configuration can be transformed into another. Through this approach, we demonstrate the practical benefits of our method.

[39] Visual manipulation of group behavior in Plecoglossus altivelis via projected dot patterns

Kohei Ohashi¹, Rei Hiraoka¹, Syoma Kamata¹, Raj Rajeshwar Malinda. Hiroaki Kawashima², Hitoshi Habe², and Takayuki Niizato^3,4

[1] Institute of Systems and Information Engineering, University of Tsukuba, Ibaraki, Japan
[2] Graduate School of Information Science, University of Hyogo, Kobe, Hyogo, Japan
[3] Faculty of Informatics, Kindai University, Higashi-Osaka, Osaka, Japan
[4] Cyber Informatics Research Institute, Kindai University, Higashi-Osaka, Osaka, Japan

Fish primarily rely on visual cues for navigation and coordination. However, the effects of temporally varying visual stimuli on group dynamics remain poorly understood. In this study, we projected rotating dot patterns—black dots (B.D.), white dots (W.D.), and a no-stimulus control—onto the bottom of an experimental tank. We then analyzed both individual and group-level behaviors of Plecoglossus altivelis. Our results showed that fish under the B.D. condition exhibited significantly stronger alignment with the direction of stimulus movement. Nearest neighbor distance (NND) analysis further revealed that group cohesion increased under the B.D. condition. In contrast, the W.D. condition led to more dispersed schooling structures. Our findings demonstrate that dynamic visual stimuli can robustly influence both individual motion and collective organization in fish. This study highlights the potential of dynamic light patterns as a non-invasive method for guiding fish schools. It also offers a novel experimental framework for exploring the visual mechanisms that govern collective motion.

[13] Improving Topological Diversity in Mutation-Based Evolving Artificial Neural Network via Multi-Objective Optimization

Kenta Okumura¹, Motoaki Hiraga², Kazuhiro Ohkura³, and Arata Masuda²

[1] Division of Mechanophysics, Kyoto Institute of Technology, Kyoto, Japan
[2] Faculty of Mechanical Engineering, Kyoto Institute of Technology, Kyoto, Japan
[3] Graduate School of Advanced Science and Engineering, Hiroshima University, Hiroshima, Japan

Neuroevolution is an approach that utilizes evolutionary algorithms to design neural networks, which can simultaneously evolve weights and neural network topology. However, neuroevolution methods often suffer from uncontrolled growth and limited diversity in network topologies. This study focuses on Mutation-Based Evolving Artificial Neural Network (MBEANN), a neuroevolution method that uses only mutation and without crossover for genetic variation. This study aims to develop MBEANN that can achieve topological diversity within the population and optimize the neural network topology while maintaining high performance. To achieve this, multi-objective optimization is incorporated into MBEANN, which explicitly encourages topological diversity by optimizing both task fitness and topological novelty. In addition, the parent selection method is changed to (μ, λ) selection to enhance the topological diversity within the population. The experimental results using Ant-v4 and Walker2d-v4 locomotion tasks demonstrate that the proposed method enhances topological diversity while maintaining performance and mitigating topological bloats.

[09] Enhancing Morphological Diversity and Demonstrating Control Robustness of Polyhedral Modular Rovers

Shingo Onizuka¹, Ryohei Michikawa², Yuto Fukao², Xixun Wang³, Haruho Mitsunaga⁴, Soshi Yoneda⁴, Yuki Takagi³, Yousuke Izuno³, Mihiro Nakabayashi⁵, Reo Nishino⁵, Yuya Shimizu⁶, Ryosuke Inoue¹, Syota Uryu¹, Akio Noda⁴, Takashi Takuma³, Hiroshi Oku³, Tomohiro Hayakawa⁷, Tetsushi Kamegawa⁶, Fumitoshi Matsuno³, and Ryusuke Fujisawa¹

[1] Graduate School & Faculty of Environmental Engineering, The University of Kitakyushu, Fukuoka, Japan
[2] Graduate School of Engineering, Kyoto University, Kyoto, Japan
[3] Graduate School & Faculty of Engineering, Osaka Institute of Technology, Osaka, Japan
[4] Graduate School & Faculty of Robotics & Design Engineering, Osaka Institute of Technology, Osaka, Japan
[5] Graduate School of Science and Engineering, University of Toyama, Toyama, Japan
[6] Graduate School of Environment, Life, and Natural Science and Technology, Okayama University, Okayama, Japan
[7] Department of Mechanical Engineering, Faculty of Engineering, Shizuoka University, Shizuoka, Japan

Modular robots can adapt to diverse missions by reconfiguring their body structures. However, highly connectable modules often compromise motion robustness. In this study, we aim to enhance the operational stability of a four-wheeled rover composed of polyhedral BODY modules with 16 symmetric docking faces, through three lightweight control improvements: an empirical PID retuning, initial posture correction, and synchronised LIMB actuation. A total of 32 field trials were conducted across three terrains, two body-weights, two speeds, and two symmetry forms. The control upgrades significantly reduced the mean standard deviation of motor current by 1.66W (95 % CI: [−3.2, −0.1] W; p = 0.037), as revealed by ordinary least squares regression controlling for body-weight, terrain, form, and velocity, with up to a 46% variance reduction under the heavy-mat condition. These results demonstrate that software-level tuning can effectively enhance the motion robustness of highly connectable hardware without the need for additional sensors or mechanical redesign.

[33] Position Estimation Experiment on Bio-inspired Robotic Mouse with Inertial Measurement Unit

Gantogoo Oyunbat¹, Sven Lange¹, Zhenshan Bing², and Florian Röhrbein¹

[1] Professorship of Neurorobotics, Faculty of Computer Science, Chemnitz University of Technology, Chemnitz, Germany
[2] Chair of Robotics, Artificial Intelligence and Real-time Systems, School of Computation, Information and Technology, Technical University of Munich, Germany

This paper presents an empirical study of the neurorobotic mouse ”TUCmouse”. The main aim of this study was to integrate an Inertial Measurement Unit (IMU) in the neurorobotic mouse for indoor position estimation. Visionbased state estimators allow robotic agents to accurately estimate their relative position. In visually degraded situations such as dark or foggy environments, these systems have difficulty getting the correct estimation, which can lead to robots colliding with objects in the environment or getting lost [1]. Proprioceptive state estimators based on the IMU can help compensate for the issue of false position estimation. Some robots on the market or custom-made robots are equipped with dedicated sensors for detecting ground contact, while some do not necessarily have them. Adding such specific sensors can be demanding, as the robot design might need to be reconsidered or changed. For indoor navigation and state estimation applications, the use of an IMU is needed, and most robotic systems already have one integrated. Although many state-of-the-art methods for dead-reckoning and indoor positioning were developed for larger quadrupeds or other robotic platforms, adapting them to a small, lightweight bioinspired robotic mouse introduces additional challenges related to scale, sensor noise, and body movements. Despite these challenges, we tested IMU-only position estimation algorithms on TUCmouse and demonstrated the potential of low-cost inertial sensing for approximate localization on small robots with bioinspired skeletal build.

[43] Modeling information propagation in robot swarms through epidemiological models

Giuseppe Antonio Patarino¹, Volker Strobel and¹, and Marco Dorigo¹

[1] IRIDIA, University Libre de Bruxelles, Bruxelles, Belgium

We present an epidemiology-inspired method to predict information dissemination in a robot swarm performing a random walk. We map the Susceptible–Infected (SI) model—a minimal two-compartment epidemic model—onto swarm communication. In our setting, the effective contact rate is empirically modeled as a function of robot density, communication range, and speed. Using ARGoS simulations, we generate dissemination curves under varying swarm parameters and fit them with a logistic model, yielding a compact predictive equation for estimating the time to achieve any desired level of information coverage. This fitted model provides a practical tool for swarm designers to select appropriate parameters—such as robot swarm density, communication range, and speed—to meet specific latency requirements in information propagation. Future work will extend this framework to other mobility patterns and to disruptive factors such as communication jamming.

[50] Modular Cross-Platform Software Design for the Autonomy Core of Decentralized Heterogeneous Mobile Robot Swarms

Alexander Puzicha¹

[1] Department of Computer Science, TU Dortmund University, 44227 Dortmund, Germany

Developing modular autonomy systems for decentralized, heterogeneous robot swarms poses significant challenges, particularly when targeting multiple processor architectures under tight resource constraints. Existing research frameworks, often based on Python or ROS, incur high runtime and energy costs, rendering them unsuitable for large-scale outdoor deployments that require compressed, resilient, low-bandwidth communication over long distances. To address these limitations, we present a novel modular, cross-platform autonomy core designed for both real and virtual agents. The system supports seamless deployment of over 70 agents without code modifications between simulation and physical platforms. Its architecture balances flexibility, scalability, and runtime efficiency, making it well-suited for deployment in diverse and constrained environments. It also supports integration with existing simulators as co-simulation or multilevel extensions. A central component of the system is a model predictive controller (MPC) based on Advanced Control Particle Belief Propagation (ACPBP). We evaluate ACPBP in a comprehensive set of experiments against state-of-the-art gradient-based optimizers. The results demonstrate its robustness and efficiency in generating feasible trajectories across various vehicle-like systems, while also revealing its limitations in stabilization tasks involving oscillator dynamics.

[51] Understanding visual attention behind bee-inspired UAV navigation

Pranav Rajbhandari¹, Abhi Veda¹, Matthew Garratt¹, Mandayam Srinivasan², and Sridhar Ravi¹

[1] School of Engineering and Technology, University of New South Wales, Canberra, Australia
[2] Queensland Brain Institute, University of Queensland, Brisbane, Australia

Bio-inspired design is often used in autonomous UAV navigation due to the capacity of biological systems for flight and obstacle avoidance despite limited sensory and computational capabilities. In particular, honeybees mainly use the sensory input of optic flow, the apparent motion of objects in their visual field, to navigate cluttered environments. In our work, we train a Reinforcement Learning agent to navigate a tunnel with obstacles using only optic flow as sensory input. We inspect the attention patterns of trained agents to determine the regions of optic flow on which they primarily base their motor decisions. We find that agents trained in this way pay most attention to regions of discontinuity in optic flow, as well as regions with large optic flow magnitude. The trained agents appear to navigate a cluttered tunnel by avoiding the obstacles that produce large optic flow, while maintaining a centered position in their environment, which resembles the behavior seen in flying insects. This pattern persists across independently trained agents, which suggests that this could be a good strategy for developing a simple explicit control law for physical UAVs.

[38] Evolutionary Approach for Optimization of Perceptual Abilities and Group Composition in Heterogeneous Robotic Swarms

Asad Razzaq¹, Senshu Wakayama¹, Mazhar Manzoor¹, and and Toshiyuki Yasuda²

[1] Graduate School of Science and Engineering, University of Toyama, Toyama, Japan
[2] Faculty of Engineering, University of Toyama, Toyama, Japan

Swarm robotic systems (SRS) comprise multiple simple robots that achieve complex tasks through local interactions and decentralized control. This study proposes an evolutionary framework for designing perceptual abilities and swarm composition in heterogeneous SRS. The system autonomously adapts to task demands by evolving sensor configurations and group composition. Experimental results in cooperative exploration tasks demonstrate that the proposed approach improves performance and flexibility compared to fixed configurations, especially in environments with varying complexity.

[35] Enhancing LLM Inference with Human Expert Knowledge: A Case Study on Mobile Robotics Fault Diagnosis and Prediction

Yongxu Ren¹, Felix Deichsel², Jürgen Seiler², Andr´e Kaup², and Philipp Beckerle¹

[1] Autonomous System and Mechatronics,Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany
[2] Multimedia Communications and Signal Processing, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany

The rapid advancement of large language models (LLMs) has revealed their impressive capabilities in semantic understanding and logical reasoning over text-based data. These developments open new avenues for integrating LLMbased inference with high-value domain knowledge from human experts. However, expert knowledge is often unstructured and limited in volume, particularly within specific industrial domains, rendering traditional fine-tuning approaches impractical. In this work, we propose a novel and efficient framework that facilitates integrating expert knowledge from a Delphi study into LLMs. We validate our framework through a case study on a predictive maintenance (PdM) use case involving mobile robotic systems, demonstrating notable inference performance improvements with human experts’ knowledge and the feasibility of leveraging expert knowledge in data-scarce environments. Implementation for this work is published in https://github.com/Prae-Flott/Knowledge4LLM

[04] Recurrent neural network trained unsupervised to discover feature combinations

Luna Rusteberg¹, Minija Tamosiunaite^2,3, and Florentin Wörgötter²

[1] Institute of Computer Science, University of Göttingen, Göttingen, Germany
[2] Institute Physics III, University of Göttingen, Göttingen, Germany
[3] Faculty of Informatics, Vytautas Magnus University, Kaunas, Lithuania

This paper analyzes learning and behavior of a recurrent neural network trained using the Annealed Linear Learning (ALL) rule, where the essence of the rule is annealing of the learning rate for a neuron as soon as that neuron has high enough output. We show that in case we provide several external inputs to a recurrent neural network and train every neuron in the network using the ALL rule, inside the network a set of neurons signaling various external input combinations emerge. This was shown using different numbers of external input lines exciting the network and for different external input line excitation patterns.

[21] Bio-Inspired Spiral Soft Arm Based on an Animal Tongue

Haruhiro Sato¹, Ku Esaki², and Kazuyuki Ito³

[1] Major in Electrical and Electronic Engineering, Graduate School of Science and Engineering, Hosei University, Tokyo, Japan
[2] TransCosmos Inc., Tokyo, Japan
[3] Department of Electrical Electronic Engineering, Hosei University, Tokyo, Japan

In recent years, soft manipulators inspired by living organisms have attracted considerable attention. These manipulators can adapt flexibly to various situations through interaction with their environment. Among them, we focused on the tongue of a giraffe, which has a structure that allows it to wrap around objects completely. In this study, we developed a soft arm that mimics the behavior of the giraffe’s tongue using a simple mechanism. We conducted experiments to evaluate its performance by changing the orientation of the arm and testing its ability to grasp various objects. We also conducted experiments in narrow spaces to confirm its operable range. As a result, the proposed soft arm was able to grasp columnar objects of various shapes and sizes more stably than a conventional rigid arm, regardless of orientation, even in confined spaces

[47] Optimization of Autonomous Transformation Paths in Modular Robots via Structure Graph Encoding and Isomorphism Determination

Kenichiro Satonaka¹, Ryusei Nishii¹, Ryota Kinjo^1,2, Seiichi Ohashi², Tomoya Negoro², Yuki Takagi¹, Hiroshi Oku^1,2, Yuki Tanigaki^1,2, Koki Harada³, Ryo Ariizumi⁴, Tomohiro Simomoura¹, Guang Yang^1,2, Xixun Wang^1,2, and and Fumitoshi Matsuno^1,2

[1] Graduate School of Engineering, Osaka Institute of Technology, Osaka, Japan
[2] Department of Electronics and Information Systems Engineering, Osaka Institute of Technology, Osaka, Japan
[3] Graduate School of Engineering, Nagoya University, Nagoya, Japan
[4] Department of Mechanical System Engineering, Tokyo University of Agriculture and Technology, Tokyo, Japan

The Moonshot R&D project, “Self-evolving AI robot system for lunar exploration and human outpost construction,” aims to realize AI modular robots that integrate advanced physical capabilities with self-evolving AI learning. Modular robots face significant challenges in achieving optimal and rapid shape transitions due to the vast number of possible configurations. To address this, we first developed a Graph Neural Network (GNN) designed to learn abstract structural features inherent to modular robots—such as functional equivalence and geometric symmetry. By capturing these structural patterns, the GNN can effectively infer promising deformation paths without exhaustive search. To enable effective training of such a GNN, we developed a novel isomorphism determination method called Canonical Robot Adaptive Graph Encode (CRAGE), which converts structural graphs into canonical strings for efficient and consistent comparison. This graph-to-string encoding makes it possible to identify structurally equivalent configurations with high accuracy and speed. Moreover, CRAGE enables the rapid and scalable generation of high-quality training data for robot structures, isomorphism classification, and deformation path planning—providing essential input for the GNN. Furthermore, to incorporate geometric physical judgments into GNN, we created a system using MoveIt! and demonstrated its effectiveness by comparing it with real-world data. Combined, CRAGE and the GNN form a unified framework for fast and scalable transition planning, contributing to the realization of autonomous self-evolving modular robots.

[36] Validating the Potential of Multi-UAV Systems to Mitigate Large-scale Wildfires Based on Historical Data and Fire Modeling

Yinan Shi¹, Georgios Tzoumas¹, Thomas Richardson¹, and Sabine Hauert¹

[1] Bristol Robotics Laborotory, University of Bristol, Bristol, United Kingdom

The increasing number of wildfires in recent years has caused dangerous hazards to the environment and humans. UAVs are considered a novel technology for wildfire firefighting scenarios. In this paper, we validate the capacity and effectiveness of multi-UAV systems to mitigate large-scale wildfires based on real-world satellite data, and determine the requirements of the desired platforms. The number of UAVs, their water capacity, and water flow rate, are key parameters to mitigate the wildfires. For a wildfire of around 11.3 km2, a system with approximately 6000 highpayload UAVs is required, while the number increases sharply with increasing fire size. Although the required number of UAVs is massive, the simulation shows that it is possible to use UAVs to mitigate large-scale wildfires.

[71] Effects of An Interaction Network Structure on the Task Allocation Dynamics Generated by the Response-Threshold Model

Masashi Shiraishi^1,2, and Hiraku Nishimori²

[1] Department of Information Sciences, Hiroshima City University, Hiroshima, Japan
[2] Meiji Institute for Advanced Study of Mathematical Sciences, Meiji University, Tokyo, Japan

Social insects like ants and bees organize tasks efficiently without central control, a phenomenon often explained by the Response Threshold Model (RTM). The RTM shows how workers self-organize task allocation based on internal thresholds rather than explicit communication. This study demonstrates that network topology strongly influences task dynamics, with different structures producing distinct outcomes. For example, 2D lattices slowed task switching, scale-free networks created persistent specialists, and small-world networks balanced flexibility and specialization. Overall, combining RTM with network theory offers deeper insights into insect societies and may guide decentralized systems like swarm robotics.

[29] Autonomous Detection and Coverage of Unknown Target Areas by Multi-Agent Systems

Jie Song¹, Yang Bai², Mikhail Svinin³, and Naoki Wakamiya¹

[1] Department of Bioinformatic Engineering, University of Osaka, Osaka, Japan
[2] Graduate School of Advanced Science and Engineering, Hiroshima University, Hiroshima, Japan
[3] Graduate School of Information Science and Engineering, Ritsumeikan University, Osaka, Japan

This paper presents a novel coverage control algorithm for multi-agent systems, where each agent has no prior knowledge of the specific region to be covered. The proposed method enables agents to autonomously detect the target area and collaboratively achieve full coverage. Once an agent detects a part of the target region within its sensor range, a dynamically constructed density function is generated to attract nearby agents. By integrating this density-driven mechanism with Centroidal Voronoi Tessellation (CVT), the agents are guided to achieve optimal spatial distribution. Additionally, Control Barrier Functions (CBFs) are employed to ensure collision avoidance and maintain non-overlapping sensor coverage, enhancing both safety and efficiency. Simulation results verify that agents can independently locate and effectively cover the target area.

[06] Development of A Bio-Inspired Anthropomorphic Finger for Dexterous Robotic Hands

Pasut Suriyasomboon¹, Rene M. Suarez Flores¹, and Sajid Nisar¹

[1] Department of Mechanical and Electrical Systems Engineering, Faculty of Engineering, Kyoto University of Advanced Science, Kyoto, Japan

This paper proposes a new mechanical mechanism for anthropomorphic robotic fingers to enhance dexterity, efficiency, and adaptability for robotic manipulation tasks with a minimal number of actuators. The proposed design uses a tendon-drive mechanism that replicates human finger posture and movement, implemented through three separate sub-mechanical systems optimized for independent control. The design offers mechanical simplicity while preserving functionality, enabling complex finger movements through coordinated actuation. We develop the finger kinematics and fabricate a five-finger robot hand. Experimental evaluation demonstrates the robotic finger’s ability to perform a wide range of grasp types, including power, precision, and intermediate grasp, while maintaining smooth, human-like motion and high adaptability to object shapes.

[55] Improving DRL Policies via Explainable Action Correction Based on Accident Cause Analysis

Tenta Suzuki¹, Mao Tobisawa¹, Clive Jancen Kawaoto¹, Kaito Kumagae¹, Yuki Itoh¹, Kenji Matsuda², Junya Hoshino³, Tomohiro Harada⁴, Jyouhei Matsuoka¹, and Kiyohiko Hattori⁵

[1] Tokyo University of Technology, 1404-1 Katakuramachi, Hachioji, Tokyo 192-0983, Japan
[2] Nitori Co., Ltd., Meguro-dori Store, 6-1-18 Shimomeguro, Meguro-ku, Tokyo 153-0064, Japan
[3] Headwaters Co., Ltd., 6-5-1 Nishi-Shinjuku, Shinjuku-ku, Tokyo 163-1304, Japan
[4] Saitama University, 6-6 Asahigaoka, Hino-shi, Tokyo 191-0065, Japan
[5] Tokyo Denki University, 5 Senju Asahi-cho, Adachi-k

In recent years, Deep Reinforcement Learning (DRL) has been applied in various fields, including autonomous driving, and its effectiveness has been extensively evaluated. While DRL efficiently explores state-action spaces and demonstrates higher performance than human-designed approaches for various tasks, it faces the challenge of generating undesirable actions that designers did not anticipate due to the black-box nature of the generated rules. To address this challenge, this study aims to achieve both interpretability and validity of output actions by proposing a rule-based correction method using the learning results of DRL. Specifically, we perform knowledge distillation of a trained model obtained through DRL using XGBoost. Next, we use SHAP to quantify the contribution of each observation feature to actions. Using these results, we construct feature vectors that combine observation information and SHAP values, and perform clustering using the k-means method. Subsequently, we extract typical accident factors by calculating the average difference between observation values and SHAP values during normal and accident conditions in each cluster, and derive correction rules. The correction is applied additively to the control output during inference. To evaluate the effectiveness of the proposed method, we conducted verification using an autonomous driving simulation. As a result, this method reduced the accident rate by up to 21% while maintaining the average lap time, which indicates the efficiency of autonomous driving.

[70] Pheromone-Reward Reinforcement Learning for Trajectory Design of Multi-Agent: Evaluation Using Lane-less Self Driving as a Case Study

Tenta Suzuki¹, Yuta Tobe², Tomohiro Harada³, Johei Matsuoka¹, and Kiyohiko Hattori³

[1] Tokyo University of Technology, 1404-1, Katakuramachi, Hachiouji-shi, Tokyo, Japan
[2] Tokyo Denki University, 5 Senju Asahi-cho, Adachi-ku, Tokyo, Japan
[3] Saitama University, 225, Shimookubo, Sakura-ward, Saitama-shi, Saitama, Japan

Deep Reinforcement Learning (DRL) is emerging as an innovative solution for real-world problems such as autonomous driving. Harada et al. [1] successfully implemented driving control in a “lane-less” environment, which does not assume predefined lanes, by using DRL. However, their research faced the challenge of high collision rates even after the learning process converged[1,3]. To address this issue, this study proposes a novel framework called Pheromone-Reward Reinforcement Learning, which integrates the concept of Ant Colony Optimization (ACO) [2] into the conventional reward system. This mechanism allows an agent (vehicle) to deposit “pheromones” along its trajectory, encouraging subsequent vehicles to follow safer paths with high pheromone concentrations. By combining the individual learning from an agent’s sensors with collective intelligence acquired via pheromones, we aim to improve learning efficiency and reduce collision rates.

[41] Development of a Health Monitoring Scheme on Mechanical Connection Motion of Modular Robots

Yuki Takagi¹, Kazuki Shibata¹, Hiroshi Oku¹, Yuki Tanigaki¹, Guang Yang¹, Xixun Wang¹, Ryohei Michikawa², and Fumitoshi Matsuno¹

[1] Department of Electronics and Information Systems Engineering, Osaka Institute of Technology, Osaka, Japan
[2] Department of Mechanical Engineering and Science, Kyoto University, Kyoto, Japan

This paper proposes a health-monitoring framework for the connection sequence between modules in a modular robot designed for lunar exploration. The target modular robot consists broadly of three types of modules: body modules, limb modules, and end-effector modules. In this study, we consider a scenario in which a limb module equipped with an active coupling mechanism, placed on a pallet, connects to an end-effector module fixed to the pallet and equipped with a passive coupling mechanism. To ensure reliable inter-module connection operations under the harsh conditions of the lunar surface, it is essential to automatically monitor the entire connection process. In particular, it is important to (I) perform precise positioning of the limb module’s end-effector using camera images of AR markers attached to the end-effector, and (II) monitor the successful completion of the mechanical coupling operation. In this study, we develop an integrated monitoring scheme that checks (I) the detection of AR markers and (II) the engagement status of the mechanical coupling mechanism. The effectiveness of the proposed scheme is validated through experiments using actual hardware.

[56] Multi-Objective Optimization of Module Configuration and Task Allocation for Modular Robots of the Lunar Exploration Project

Yuki Tanigaki¹, and Yano Mikito¹

[1] Department of Electronics and Information Systems Engineering, Osaka Institute of Technology, Osaka, Japan

In task allocation for general-purpose robots, the performance of robots is typically predefined. However, in the case of modular robots, execution efficiency can vary significantly depending on the robot’s configuration. This variability makes it challenging to simultaneously optimize both the module combinations and the task allocation. To address this issue, this paper proposes a two-stage bottom-up optimization framework. In the first stage, we perform multi-objective optimization using a simplified evaluation metric that can be computed without simulations, aiming to identify Pareto-optimal robot configurations. In the second stage, we apply task scheduling optimization to each of these configurations.

[53] Efficient and Safe Bidirectional Traffic Control in Lane-less Environments Using Global Reward and Curriculum Learning

Mao Tobisawa¹, Tenta Suzuki¹, Clive Jancen Kawaoto¹, Yuki Itoh¹, Kaito Kumagae¹, Tomohiro Harada², Johei Matsuoka¹, and and Kiyohiko Hattori³

[1] Tokyo University of Technology, 1404-1, Katakuramachi, Hachiouji-shi, Tokyo, Japan
[2] Saitama University, 225, Shimookubo, Sakura-ward, Saitama-shi, Saitama, Japan
[3] Tokyo Denki University, 5 Senju Asahi-cho, Adachi-ku

Recent advances in machine learning technology have accelerated research and development in autonomous driving. Since conventional road environments employ lane-separated structures, inefficient utilization of road resources due to imbalanced traffic demand has become a critical issue. To address this problem, bidirectional traffic control in lane-less environments that leverages high-precision sensing data shared through vehicle-to-vehicle communication has attracted attention. In our previous study, we achieved bidirectional traffic in lane-less environments using deep reinforcement learning and accomplished significant improvements in road utilization efficiency. However, a safety issue arose due to an increase in vehicle collisions. Simply strengthening collision penalties resulted in vehicles learning only to follow preceding vehicles without overtaking behaviors, thus losing the advantages of lane-less environments. In this study, we propose a novel method combining global reward and curriculum learning to address the trade-off between efficiency and safety. The global reward enables the transition from individual optimization to system-wide optimization, while curriculum learning achieves gradual safety improvements. Experimental results confirmed that introducing global reward increased the evaluation tile passage count to 145%, and curriculum learning enabled gradual accident rate reduction while maintaining cooperative control capabilities. Furthermore, we identified that speed differences between vehicles are the primary cause of accidents and demonstrated that fixing vehicle speeds can significantly reduce accidents.

[37] Feasible Learning Problems for Networks of Conditioning Logical Devices: Characterization by Output Controllability with Periodic State Preservation

Tomoki Tokiwa¹, Shun-ichi Azuma¹, and Ikumi Banno¹

[1] Graduate School of Informatics, Kyoto University, Kyoto, Japan

Classical conditioning is a learning process in which an organism acquires conditioned reflexes through the repeated association of specific stimuli. Building on a recently developed mathematical model of this phenomenon, the learning problem for networks composed of such models has also been addressed. This paper generalizes the foundational model and investigates the conditions that determine the feasibility of this learning problem.

[65] Gait stability landscape of a quadruped robot governed by neuromechanical resonance

Sota Tsuchida¹, Tomoya Kamimura¹, Takatoshi Hashimoto¹, Tomoki Nakagawa¹, Yuichi Ambe², Mau Adachi¹, and and Shinya Aoi¹

We developed a quadruped robot consisting of a single body and four legs. It has been suggested that CPGs consist of two hierarchical networks composed of a rhythm generator (RG) layer and a pattern formation (PF) layer. The RG layer produces rhythmic signals for locomotion, while the PF layer translates these signals into motor commands based on limb kinematics. In this study, we developed a controller that regulates the length and orientation of each leg axis using linear and torsional springs, driven by four two-level CPGs based on our previous study [4]. Each RG layer contains a phase oscillator whose dynamics are determined by frequency, inter-CPG coupling, and sensory feedback. The PF layer determines the equilibrium length and angle of the leg springs using the oscillator phases. When we focus on walk, trot and pace, these gait patterns emerge as a result of a single phase difference between ipsilateral oscillators.

[10] Modeling Fish Schooling Behavior for Selective Fishing in Set Net Fisheries

Tomoki Umezaki¹, Masaki Ogura², Takayuki Nakamura³, Shodai Suzuki⁴, Tsutomu Takagi⁴, Tomonori Yoshikawa⁵, Shunsuke Torisawa⁵, Keitaro Kato⁵, Youhei Washio⁵, Shukei Masuma⁵, and Naoki Wakamiya⁶

[1] Graduate School of Information Science and Technology, The University of Osaka, Osaka, Japan
[2] Advanced Science and Engineering, Hiroshima University, Hiroshima, Japan
[3] Graduate School of Environmental Science, Hokkaido University, Hokkaido, Japan, [4] Department of Aquaculture Life Science, Hokkaido University, Hokkaido, Japan
[5] Department of Fisheries, Kindai University, Wakayama, Japan
[6] Graduate School of Information Science and Technology, The University of Osaka, Osaka, Japan

Fish schooling behavior has long been studied to understand collective motion in biological systems and to apply this knowledge to fields such as robotics and traffic modeling. In this paper, we propose a simplified and practical mathematical model of fish schooling behavior considering future application to selective fishing in set net fisheries. The proposed model is constructed based on three-dimensional trajectory data of Trachurus japonicus in an experimental tank, and it captures essential features of schooling with minimal parameters and limited information—focusing on only one neighbor to adapt the relative position and the swimming speed of a focal fish. Comparative evaluation with a modified Gautrais model and experimental fish movement data demonstrates that our model can reproduce the proximity of fishes, speed variability, and general alignment tendencies. The simplicity of the model makes it suitable for practical applications of selective fishing in set net fisheries where it is difficult to obtain precise data of a variety of fishes showing various behaviors.

[11] Agricultural weed control by a swarm of robots

Glenn R. Varhaug¹, Jarle Dorum², Jan Tommy Gravdahl¹, Elias Prytz², and and Henrik Vasbotten²

[1] Department of Engineering Cybernetics, Norwegian University of Science and Technology, Trondheim, Norway
[2] Kilter AS, Berghagan 3, 1405 Langhus, Viken, Norway

Herbicide application in agriculture is an effective means for combating weeds. The weed control problem can be solved with a fully autonomous precision spraying robot. In order to decrease the time spent on herbicide application, a swarm robotic solution using the AX-1 robot has been developed. The solution involves field decomposition, workload distribution, and path planning. Results show that the effect of increasing the number of robots in the swarm reduce the time spent in the field for all fields tested, regardless of size and complexity. The results do however point to distance alone being insufficient as a sole criterion for workload distribution, prompting future research to investigate how to gather a weed density distribution before performing workload distribution.

[12] From Pheromones to Policies: Reinforcement Learning for Engineered Biological Swarms

Aymeric Vellinger¹

Swarm intelligence emerges from decentralised interactions among simple agents, enabling collective problem-solving. This study establishes a theoretical equivalence between pheromone-mediated aggregation in C. elegans and reinforcement learning (RL), demonstrating how stigmergic signals function as distributed reward mechanisms. We model engineered nematode swarms performing foraging tasks, showing that pheromone dynamics mathematically mirror cross-learning updates, a fundamental RL algorithm. Experimental validation with data from literature confirms that our model accurately replicates empirical C. elegans foraging patterns under static conditions. In dynamic environments, persistent pheromone trails create positive feedback loops that hinder adaptation by locking swarms into obsolete choices. Through computational experiments in multi-armed bandit scenarios, we reveal that introducing a minority of exploratory agents insensitive to pheromones restores collective plasticity, enabling rapid task switching. This behavioural heterogeneity balances exploration-exploitation trade-offs, implementing swarm-level extinction of outdated strategies. Our results demonstrate that stigmergic systems inherently encode distributed RL processes, where environmental signals act as external memory for collective credit assignment. By bridging synthetic biology with swarm robotics, this work advances programmable living systems capable of resilient decision-making in volatile environments.

[14] Robust UAV Formation Control via Improved ESO and SMC: From Wind-Disturbed Simulations to Real-World Flights

Wei Wang¹, and Qi Wang²

[1] Autonomous, Intelligent, and Swarm Control Research Unit, Fukushima Institute for Research, Education and Innovation (F-REI), 6-1 Yazawa-machi, Gongendo, Namie Town, Futaba County, Fukushima 979-1521, Japan
[2] School of Electrical and Electronic Engineering, Nanjing Xiaozhuang University, Nanjing, China

A formation control approach for Unmanned Aerial Vehicles (UAVs) is introduced to improve the track ability when flying in environments with interference. The method is built around three main steps. First, the desired trajectory is adjusted based on UAV features to enhance trajectory information, to helps the UAVs follow the path more steadily. Second, a improved extended state observer (IESO) is designed to better estimate disturbances. Third, a distributed controller is designed using a type of sliding mode control (SMC), which takes both the improved trajectory and disturbance feedback into account to achieve more accurate and stable formation tracking. The effectiveness of this method is verified through both simulation and real-world experiments, offering a practical improvement for precise and stable UAV formation flight under uncertain conditions.

[23] Autonomous Exploration of Mobile Robots in Unknown Uneven Terrain Inspired by Insect Locomotion Strategies

Xixun Wang¹, Ryohei Michikawa², and Fumitoshi Matsuno¹

[1] Osaka Institute of Technology, Asahi, Osaka, Japan
[2] Kyoto University, Nishikyo, Kyoto, Japan

This study aims to create a navigation method to enable mobile robots to explore unknown environments with elevation differences, even when self-localization reliability is low and no predefined waypoints are available. Inspired by insects’ pheromone-based navigation, we propose a method that constructs a local digital pheromone map recording pheromone trails and uses it to generate local motion. To verify the effectiveness of the proposed algorithm, we conducted experiments using the autonomous flipper-equipped robot FUHGA4 at RoboCup Japan Open 2025.

[34] A Human-in-the-Loop DAG Refinement Framework for LLM-Driven Multi-Robot Construction Task Planning

Yongdong Wang¹, Runze Xiao², Jun Younes Louhi Kasahara²

[1] The University of Tokyo, Tokyo, Japan
[2] Kindai University, Hiroshima, Japan
[3] University of Tsukuba, Ibaraki, Japan

We propose a human-in-the-loop refinement framework based on Directed Acyclic Graphs (DAGs) for Large Language Model (LLM)-driven multi-robot construction task planning. The framework incorporates a Question- Answering (QA) LLM module to translate the operator’s natural language instructions into a structured DAG representing atomic task dependencies. A Task Interface module presents the generated task graph to the operator for inspection and feedback. Through interactive dialogue, the operator can iteratively refine the task plan and intervene to prevent the execution of potentially unsafe actions. Experimental results demonstrate that the framework enables real-time responsiveness and effectively blocks incorrect DAG task plans from being dispatched to the robot team. In error refinement evaluations using a public dataset, the Llama3.1 8B model required an average of 1.42 dialogue turns to correct task dependency errors, whereas the GPT-4.1 model achieved error correction in just 1.17 turns. Through the iterative refinement mechanism, the framework is capable of transforming initially erroneous subtasks into safe and executable plans, enabling the reliable deployment of LLMs in multi-robot construction sites.

[32] Simultaneous optimization of structure and control of modular robots on rough terrain considering leg and wheel modules

Xixun Wang¹, Tomohiro Shimomura¹, Ryo Ariizumi², and Fumitoshi Matsuno¹

[1] Osaka Institute of Technology, Asahi, Osaka, Japan
[2] Tokyo University of Agriculture and Technology, Koganei, Tokyo, Japan

This study aims to find the best configuration of a modular robot that is able to exchange the equipping modules for various uneven terrains. Legged robots offer superior traversal capabilities in rough environments, while wheeled robots provide higher efficiency on flat surfaces. Therefore, designing a robot adaptable to various terrains requires consideration of both control strategies and structural configurations. To address this challenge, we employ machine learning techniques to simultaneously optimize the robot’s structure and control parameters. While previous research has focused primarily on legged robots, legs are not always the most efficient means of locomotion on relatively flat terrain. To expand the design space, we introduce wheel modules alongside conventional leg structures. Three types of terrain are defined to evaluate the proposed system, and we validate the effectiveness of a hybrid wheel-leg robot in navigating each environment.

[16] Enhancing Swarm Foraging Efficiency via Dynamic Task Switching and Threshold Diversity

Taiki Watanabe¹, Yuki Tanigaki², Tomohiro Hayakawa³, Fumitoshi Matsuno², and and Ryusuke Fujisawa⁴

[1] Department of Human Intelligence Systems,Kyushu Institute of Technology, Fukuoka, Japan
[2] Department of Electronics and Information Systems Engineering, Osaka Institute of Technology, Osaka, Japan
[3] Department of Mechanical Engineering, Shizuoka University, Shizuoka, Japan, [4] Department of Information Systems Engineering, The University of Kitakyushu, Fukuoka, Japan

This study addresses the efficiency of swarm robot foraging in an environment with multiple food patches, each containing multiple food items. To this end, we propose an evaluation metric called ”Foraging Utility,” which integrates the number of food items, distance, and temporal factors. Furthermore, based on this metric, we developed a distributed algorithm in which robots autonomously switch among three tasks: ”Searcher,” ”Picker,” and ”Carrier”. In simulation evaluations, we first confirmed the effectiveness of a model where all robots share a common threshold. We then demonstrated that a model assigning diverse thresholds to individual robots using an asymmetric Laplace distribution suppresses congestion and significantly improves foraging efficiency. In conclusion, we show that the proposed combination of dynamic task allocation and threshold diversity maximizes collection efficiency, particularly in medium to large-scale environments.

[62] Utilization of Cohesion and Adhesion Forces in Plasmodial Slime Molds for Network Information Transfer

Kristina Wogatai¹, and Wilfried Elmenreich¹

[1] Department of Networked and Embedded Systems, University of Klagenfurt, Klagenfurt, Austria

Cohesion and adhesion forces are crucial for understanding the behaviour of plasmodial slime molds, such as Physarum polycephalum. Despite their unicellular nature, these organisms exhibit fluid-like dispersal and complex information processing. This study investigates how these forces influence slime mold locomotion, nutrient acquisition, and network formation, thereby highlighting their role in collective behaviour and communication. The dispersal and connectivity of slime mold pseudopodia are simulated using the SISMO-Py simulation tool under different spatial distributions of food sources and hatching probabilities. The results show that closely spaced food sources encourage the formation of cohesive, dense networks that enable efficient local communication. In contrast, widely spaced sources activate adhesiondriven dispersal, resulting in extended, less redundant but more flexible networks. These findings emphasise the adaptive strategies employed by slime molds in dynamic environments, suggesting potential applications in bio-inspired network design and information transmission.

[45] AutoPercep: A Pipeline for Onboard Neighbor Position Estimation Toward Scalable Swarm Robotics

Ruiheng Wu¹, Oliver Deussen^1,3, Iain D. Couzin^2,3,4, and Liang Li^2,3,4

[1] Department of Computer and Information Science, University of Konstanz, Germany
[2] Department of Collective Behaviour, Max Planck Institute of Animal Behavior, Konstanz, Germany
[3] Centre for the Advanced Study of Collective Behaviour, University of Konstanz, Konstanz, Germany
[4] Department of Biology, University of Konstanz, Konstanz, Germany

Swarm robotics, inspired by the collective behaviors found in nature, plays a vital role in enhancing scalability, resilience, and efficiency in multi-robot systems. A major challenge in deploying swarm robotics on a large scale is enabling robots to determine their neighbors’ relative positions using onboard sensors, mimicking the spatial awareness observed in animal groups. Current onboard methods, such as SLAM, require intensive computation and are not ideal for decoding neighbor information. Artificial neural networks are effective when trained on high-quality, large datasets, which are often the primary bottleneck. In this paper, we propose a general pipeline that simplifies the collection of large training datasets and enables detection with lightweight neural networks to estimate the neighbor’s relative position with minimum manual work. As a quick demonstration with two robots equipped with an onboard camera and a Raspberry Pi 4B, we use a motion capture system to label the neighbor’s relative position, synchronized with the onboard camera. By collecting over 10,000 images in just 90 minutes, we successfully train and compare three neural networks to estimate the neighbor’s relative position. Our pipeline highlights the feasibility of onboard perception and control with scalable large swarm robotics, offering promising prospects for deployment beyond the lab environment.

[31] Vision-Based Docking of Limited-DoF Modules for Reconfigurable Modular Robots

Guang Yang¹, Xixun Wang¹, Ryohei Michikawa², Kiona Hosotani², Yuya Shimizu³, Tetsushi Kamegawa³, Ryusuke Fujisawa⁴, and Fumitoshi Matsuno¹

[1] Osaka Institute of Technology, Osaka, Japan
[2] Kyoto University, Kyoto, Japan
[3] Okayama University, Okayama, Japan
[4] The University of Kitakyushu, Kitakyushu, Japan

Self-reconfigurable modular robots offer high modularity and structural adaptability by employing low-degree-of-freedom (low-DoF) modules, enabling lightweight and simplified mechanical designs. However, the limited number of degrees of freedom presents challenges in achieving precise module docking. This paper proposes a high-accuracy docking method that utilizes camera-based relative pose estimation and visual feedback control, taking into account the initial positional and orientational configuration of the modules. The proposed approach enables autonomous alignment and connection of modules despite kinematic constraints. Experimental validation using a physical prototype demonstrates the effectiveness and feasibility of the method in real-world settings.

[67] Wavy synchronization in locomotion of train millipedes : From automatic tracking to a phase oscillator model

Momiji Yoshikawa¹, and Ikkyu Aihara²

[1] Department of Mechanical Science, Graduate School of Engineering Science, The University of Osaka, Osaka, Japan
[2] Graduate School of Advanced Science and Engineering, Hiroshima University, Hiroshima, Japan

We studied the wavy synchronization in the locomotion of train millipedes (Parafontaria laminata armigera) both experimentally and theoretically. First, we recorded the walking behavior of the millipedes and used automated tracking software (DeepLabCut) to quantify the phase dynamics of each leg. The analysis revealed a stable wavy synchronization in the walking millipedes. Second, we extended the framework of phase oscillator model by assuming asymmetric interactions to account for forward locomotion. Finally, numerical simulations with estimated parameters demonstrated that this model successfully reproduces the wavy synchronization observed in the experiments. Important future directions include investigating walking adaptation on different terrains and applying locomotion of the millipedes to multi-legged bio-inspired robots.

[46] Development of a Bio-Inspired Takeoff Mechanism for Flapping-Wing Micro Aerial Vehicles

Mohamed Aziz Zouita¹, Ahmad Hammad¹, and Sophie F. Armanini²

[1] Chair of e-Avation, Technical University of Munich, Ottobrunn, Germany
[2] Department of Aeronautics, Imperial College London, London, UK

Flapping-wing micro aerial vehicles (FWMAVs) are cutting-edge bio-inspired MAVs mimicking the flight of natural fliers. These vehicles show promising applications; however, they are constrained due to being hand-launched for takeoff, which limits their capability, especially when taking off from a natural habitat. The energy requirement for takeoff of such FWMAVs poses significant challenges, especially for autonomous ground-based takeoff. Drawing inspiration from the crouch-and-jump mechanism used by birds, this study presents an underactuated, bio-inspired takeoff system that employs rolling joints and tendon-like elastic elements to store and release energy efficiently. To approach near-minimum energy takeoff conditions, the minimum lift and velocity required for flight were analytically determined and used to formulate a constrained optimization problem. A novel mechanical design was developed and simulated, featuring simplified bird-like leg segments with rolling joints and strategically positioned springs that emulate tendon extension and recoil. Actuation is achieved through a single motor that winds and releases a rope, enabling a compact and lightweight mechanism while maintaining biomechanical fidelity. This work proposes a lightweight and energy-efficient launch strategy for FWMAVs, leveraging principles from avian biomechanics and mechanical simplification.