Periodic Reporting for period 4 - M-Runners (Modal Nonlinear Resonance for Efficient and Versatile Legged Locomotion)
Periodo di rendicontazione: 2023-12-01 al 2024-11-30
The aim of M-Runners is to thoroughly advance the understanding of fundamental dynamic principles of legged locomotion to the point that those principles can be used to design robots which display similar motion characteristics, versatility, and efficiency as their biological paragons. The central hypothesis of the project is that biological locomotion is fundamentally determined by the mechanical resonance properties of the body and that a breakthrough in robot locomotion is essentially linked to understanding and exploiting these phenomena. If body design is such that walking and running correspond to intrinsic periodic motions of the body, then the control is simple and efficiency and robustness are natural consequences. However, large amplitude nonlinear oscillations of such complex systems were theoretically not well understood before the euROBIN project. Mathematical methods to describe, analyze, design and control elastic resonant robots were lacking to a large extent. The project was thus dedicated to develop a new theory of nonlinear oscillations, applicable to elastic multibody systems, be they biologic or robotic.
M-Runners performed interdisciplinary research at the border between robotics, nonlinear dynamical systems, biomechanics, and machine learning. We take inspiration from biology regarding the basic motion sequences and the muscle arrangements (couplings, redundancies, compliance distributions). Conversely, we expect our theory to generate new hypotheses for a deeper understanding of locomotion biomechanics and its control by the nervous system.
The project succeeded to develop a new theory of nonlinear oscillations, applicable to elastic multibody systems, in particular to quadruped robots with muscle-like, elastic actuation. Several such robots with increasing complexity and versatility were developed in a rapid prototyping approach, as validation platforms for the developed methodology.
We could demonstrate that most of the gait patterns known from biological quadrupeds, such as walking, trotting, or bounding can be efficiently realized on the robot based on the new nonlinear resonance theory and the distinct classes of intrinsic body oscillations discovered with it. For the first time various gaits based on elastic resonant locomotion have been realized on full quadruped robots. Moreover, we have demonstrated that the intrinsic, natural motions of elastic robots can also be easily discovered directly on the hardware by machine learning algorithms.
The quadruped robots were used in space experiments, in which, for the first time, astronauts commanded such robots from the international Space Station ISS in planetary exploration scenarios.
Applications of the technology reach, however, from health-care over personal-assistance to disaster management.
M-Runners performed interdisciplinary research at the border between robotics, nonlinear dynamical systems, biomechanics, and machine learning. We take inspiration from biology regarding the basic motion sequences and the muscle arrangements (couplings, redundancies, compliance distributions). Conversely, we expect our theory to generate new hypotheses for a deeper understanding of locomotion biomechanics and its control by the nervous system.
The project succeeded to develop a new theory of nonlinear oscillations, applicable to elastic multibody systems, in particular to quadruped robots with muscle-like, elastic actuation. Several such robots with increasing complexity and versatility were developed in a rapid prototyping approach, as validation platforms for the developed methodology.
We could demonstrate that most of the gait patterns known from biological quadrupeds, such as walking, trotting, or bounding can be efficiently realized on the robot based on the new nonlinear resonance theory and the distinct classes of intrinsic body oscillations discovered with it. For the first time various gaits based on elastic resonant locomotion have been realized on full quadruped robots. Moreover, we have demonstrated that the intrinsic, natural motions of elastic robots can also be easily discovered directly on the hardware by machine learning algorithms.
The quadruped robots were used in space experiments, in which, for the first time, astronauts commanded such robots from the international Space Station ISS in planetary exploration scenarios.
Applications of the technology reach, however, from health-care over personal-assistance to disaster management.
The fundamental contribution of the M-Runner project is a theory of nonlinear oscillation modes for general Riemannian Euler-Lagrange systems, which is applicable to such complex systems as robots as well as for better understanding motions of human and animal bodies. The project had the very ambitious goal of building up a theoretical fundament which goes massively beyond the standard theory used in robotics and in general in engineering for modelling the oscillation and resonance properties of elastic multi-body systems. This goal was reached to a very large extent and is certainly the major methodical achievement of the project.
Based on concepts from mathematical mechanics (Hamiltonian Systems) and differential geometry (Riemannian geometry, algebraic topology) we found out that indeed one can expect the considered systems to have at least as many modal oscillations as their linearized model has. We reviewed and systematized the results from over one century of literature in mathematics, physics and engineering in this field and added several new definitions and theorems to be able to address the complex systems of interest. This research resulted in an extensive review paper in the Annual Reviews in Control and was presented in a plenary talk at the World Congress of the International Federation of Automatic Control. Several papers on these fundamentals and on their application for quadruped locomotion were published. Moreover, a software toolbox was developed, which identifies the nonlinear modes for systems given their dynamics model and starting from the normal modes of the linearized model.
Based on the theory above, a new generation of elastic quadruped robots has been developed, which were used for validating locomotion gaits know from biology such as walking, trotting, bounding. We demonstrated that those gaits are strongly related to and facilitated by the elastic resonance properties of the body discovered with our theoretical framework.
In order to generate the elastic quadruped locomotion on the prototypes, both model-based approaches and machine learning (reinforcement learning) methodologies were developed and implemented. The reinforcement learning open-source software StableBaseline3 developed by one of the PhD students of the project received more than 2700 citations in three years,
The robot Bert was used in space experiments involving astronauts from the International Space Station ISS. Astronauts directly made use of the various motion capabilities of the robot commanding it to acquire soil samples in a space exploration scenario. The experiment was featured in an extensive article in the GEO magazine.
Finally, our theory generated new insights for a deeper understanding of locomotion biomechanics and its control by the nervous system. We have shown psychophysical experiments than humans can easily excite complex nonlinear oscillations, even for systems they have never experienced before. Moreover, they can adapt within seconds if dynamic properties such as frequency or oscillation shape are modified. Moreover, we have demonstrated that humans can adapt their muscle coordination easily to exploit resonance properties for efficiency of motion.
Main dissemination results:
- 45 publications in leading journals and conferences in Robotics, control, neuroscience
- 6 patents
- ERC PoC Grant SwingBots was granted for validation of the project technology
- Experiments conducted with the quadruped robots from the International Space Station ISS, described in articles in the GEO Magazine and featured in German and French National television.
- Participation at international fairs and exhibitions such as Festival of the Future (Festival der Zukunft) 2023 in Munich, CLAWAR 2024 in Kaiserslautern and Humanoids 2024 in Nancy.
- Press releases at TUM and DLR, wide media coverage of ISS experiments and quadruped robots.
- StabeBaseline 3 Reinforcement Learning Software has more than 2700 citations on google scholar, and more than 1700 forks and 9600 stars on GitHub
Based on concepts from mathematical mechanics (Hamiltonian Systems) and differential geometry (Riemannian geometry, algebraic topology) we found out that indeed one can expect the considered systems to have at least as many modal oscillations as their linearized model has. We reviewed and systematized the results from over one century of literature in mathematics, physics and engineering in this field and added several new definitions and theorems to be able to address the complex systems of interest. This research resulted in an extensive review paper in the Annual Reviews in Control and was presented in a plenary talk at the World Congress of the International Federation of Automatic Control. Several papers on these fundamentals and on their application for quadruped locomotion were published. Moreover, a software toolbox was developed, which identifies the nonlinear modes for systems given their dynamics model and starting from the normal modes of the linearized model.
Based on the theory above, a new generation of elastic quadruped robots has been developed, which were used for validating locomotion gaits know from biology such as walking, trotting, bounding. We demonstrated that those gaits are strongly related to and facilitated by the elastic resonance properties of the body discovered with our theoretical framework.
In order to generate the elastic quadruped locomotion on the prototypes, both model-based approaches and machine learning (reinforcement learning) methodologies were developed and implemented. The reinforcement learning open-source software StableBaseline3 developed by one of the PhD students of the project received more than 2700 citations in three years,
The robot Bert was used in space experiments involving astronauts from the International Space Station ISS. Astronauts directly made use of the various motion capabilities of the robot commanding it to acquire soil samples in a space exploration scenario. The experiment was featured in an extensive article in the GEO magazine.
Finally, our theory generated new insights for a deeper understanding of locomotion biomechanics and its control by the nervous system. We have shown psychophysical experiments than humans can easily excite complex nonlinear oscillations, even for systems they have never experienced before. Moreover, they can adapt within seconds if dynamic properties such as frequency or oscillation shape are modified. Moreover, we have demonstrated that humans can adapt their muscle coordination easily to exploit resonance properties for efficiency of motion.
Main dissemination results:
- 45 publications in leading journals and conferences in Robotics, control, neuroscience
- 6 patents
- ERC PoC Grant SwingBots was granted for validation of the project technology
- Experiments conducted with the quadruped robots from the International Space Station ISS, described in articles in the GEO Magazine and featured in German and French National television.
- Participation at international fairs and exhibitions such as Festival of the Future (Festival der Zukunft) 2023 in Munich, CLAWAR 2024 in Kaiserslautern and Humanoids 2024 in Nancy.
- Press releases at TUM and DLR, wide media coverage of ISS experiments and quadruped robots.
- StabeBaseline 3 Reinforcement Learning Software has more than 2700 citations on google scholar, and more than 1700 forks and 9600 stars on GitHub
The major steps beyond the state of the art were:
- A new theory on nonlinear oscillations of elastic multi-body systems
- Elastic quadruped robot prototypes for validation of elastic locomotion, including known gaits from biology such a walking, trotting and bounding
- Control methods based on the nonlinear oscillation theory as well as on reinforcement learning for elastic locomotion of quadruped robots
- First experiments from the international space station with quadruped robots being steered by astronauts in planetary exploration scenarios
- A new theory on nonlinear oscillations of elastic multi-body systems
- Elastic quadruped robot prototypes for validation of elastic locomotion, including known gaits from biology such a walking, trotting and bounding
- Control methods based on the nonlinear oscillation theory as well as on reinforcement learning for elastic locomotion of quadruped robots
- First experiments from the international space station with quadruped robots being steered by astronauts in planetary exploration scenarios