Deep reinforcement learning for modeling human locomotion control in neuromechanical simulation.
Seungmoon Song1, Łukasz Kidziński2, Xue Bin Peng3
1Department of Mechanical Engineering, Stanford University, Stanford, CA, USA. smsong@stanford.edu.
Journal of Neuroengineering and Rehabilitation
|August 17, 2021
Summary
Deep reinforcement learning advances human motor control modeling in neuromechanical simulations. A competition accelerated novel movement generation, paving the way for biomechanics and rehabilitation research.
Area of Science:
- Biomechanics and Motor Control
- Computational Neuroscience
- Robotics
Background:
- Modeling human movement in novel environments is a significant challenge.
- Current neuromechanical simulations excel at basic locomotion but struggle with higher-level control.
- Deep reinforcement learning (RL) offers potential for complex human movement modeling.
Purpose of the Study:
- To review the application of RL in neuromechanical simulations for human motor control.
- To present the "Learn to Move" competition and software platform.
- To accelerate interdisciplinary collaboration in human movement simulation.
Main Methods:
- Review of neuromechanical simulation techniques.
- Fundamentals of reinforcement learning applied to human locomotion.
- Analysis of results from the "Learn to Move" competition (2017-2019).
Main Results:
- Top teams utilized advanced deep RL to achieve unprecedented motions.
- Novel movements like rapid turning and walk-to-stand transitions were demonstrated.
- The competition fostered innovation in RL for neuromechanical simulations.
Conclusions:
- Deep RL is a powerful tool for modeling complex human motor control.
- The "Learn to Move" initiative successfully spurred advancements in the field.
- Future work will extend the competition to further biomechanics and rehabilitation research.
