Learning Athletic, Context-Adaptive Robot Locomotion
Interdisciplinary Research Achievement �This project advances legged robot locomotion by developing two approaches: Multiplicity of Behavior (MoB) for rapid adaptation to diverse tasks and environments, and DribbleBot, a system for dexterous ball manipulation under real-world conditions. MoB enables real-time strategy selection, while DribbleBot demonstrates dynamic control through simultaneous locomotion and manipulation. Both approaches leverage reinforcement learning and address challenges of varying environment and task contexts for dynamic, contact-rich motor skills. The project team also organized the workshop on Sim-to-Real Robot Learning: Locomotion and Beyond.
Impact on Artificial Intelligence�This project uses reinforcement learning to address challenges in legged robot locomotion. This approach consists of two primary components. Multiplicity of Behavior (MoB) proposes learning a single policy that encodes a structured family of locomotion strategies, which can solve training tasks in various ways. This approach allows for real-time strategy selection and rapid adaptation to new tasks or environments without time-consuming retraining, demonstrating a versatile and efficient AI solution for legged robots. The DribbleBot system showcases a novel application of AI in dexterous ball manipulation under real-world conditions, using reinforcement learning to train policies in simulation and transfer them to physical robots. DribbleBot overcomes critical challenges of accounting for variable ball motion dynamics on different terrains and perceiving the ball using body-mounted cameras under onboard computing constraints.
Impact on Fundamental Interactions
This project incorporates several inductive biases from physics, which contribute to its effectiveness in developing advanced locomotion policies for legged robots; namely: structured control policies, generalization, dynamics and kinematics, environmental interactions in physical systems, and sensory observations.
Pulkit Agrawal, Gabriel B Margolis, Yandong Ji (MIT)
The NSF Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) is �supported by National Science Foundation under Cooperative Agreement PHY-2019786
Video: Like a human athlete, DribbleBot operates from onboard perception and dynamically controls a ball across a wide variety of natural terrains including grass, mud, snow, and pavement.
Outlook & References
The next steps will be to extend the perception capabilities of the system to automatically adapt to different contexts by taking environment geometry and semantic appearance into account.
[1] https://github.com/Improbable-AI/walk-these-ways; [2] https://arxiv.org/abs/2212.03238 [3] https://arxiv.org/abs/2304.01159