Overview
Human movement depends on complex interactions between the brain, nervous system, muscles, and body. The brain does not calculate the exact force required for every muscle and joint; instead, the body’s mechanical properties and neural reflexes help coordinate movement automatically.
Led by Daniel Häufle, Georg Martius, and Pierre Schumacher, this project investigated whether morphological computation could help AI systems learn human-like movement more efficiently. Morphological computation refers to the way the body itself contributes to movement through its structure, materials, and passive mechanical responses.
The project was carried out at the University of Tübingen between 2021 and 2024.
Key results
- Developed learning methods for controlling computer models with many muscles and degrees of freedom.
- Showed that large musculoskeletal systems are difficult to train, but that carefully guided learning can produce stable, human-like movements.
- Generated bipedal walking patterns that remained stable on uneven terrain.
- Demonstrated that simple muscle-like properties can make learned movements more natural and robust.
- Combined reinforcement learning with biologically plausible objectives, rather than training AI systems only to minimise mathematical error.
- Transferred a walking pattern learned in simulation to a real robot without additional training, demonstrating a promising approach to sim-to-real transfer.
Why it matters
The project shows how the physical structure of a body can contribute to intelligent movement and reduce the burden on a control system. This could help researchers build robots that move more efficiently, adapt more naturally to disturbances, and interact more safely with their surroundings. The findings may support future applications in healthcare, rehabilitation, prosthetics, and advanced manufacturing.
This project produced the following peer-reviewed publications:
- Learning to Control Emulated Muscles in Real Robots: A Software Test Bed for Bio-Inspired Actuators in Hardware (BioRob 2024): https://doi.org/10.1109/BioRob60516.2024.10719699
- DEP-RL: Embodied Exploration for Reinforcement Learning in Overactuated and Musculoskeletal Systems (ICLR 2023): https://openreview.net/forum?id=C-xa_D3oTj6
- Emergence of natural and robust bipedal walking by learning from biologically plausible objectives (iScience 2025): https://doi.org/10.1016/j.isci.2025.112203
- Learning with Muscles: Benefits for Data-Efficiency and Robustness in Anthropomorphic Tasks (CoRL 2022): https://openreview.net/forum?id=Xo3eOibXCQ8
