Reproduce “Learning Force Control for Legged Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tifanny Portela, Gabriel B. Margolis, Yandong Ji, Pulkit Agrawal
Researchers at the Improbable AI Lab at MIT developed a method to train reinforcement learning policies for direct whole-body force control in legged manipulators, eliminating the need for force-torque sensors. The system achieved real-world end-effector force tracking with 5-10 N error and allowed the robot to apply pulling forces up to 90 N, significantly extending its manipulation capabilities and enabling compliant interaction.
Key insights
- Direct, user-commandable force control can be effectively learned for whole-body legged manipulators using only proprioceptive sensing and simulated external force interactions.
- A unified policy enables seamless coordination of the robot's legs and arm, allowing for efficient force application, expanded workspace, and compliant behavior.
- The learned force control inherently supports capabilities like impedance control, gravity compensation, and robust compliance to disturbances, enhancing human-robot interaction and kinesthetic teaching.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.