Ep#49: Learning a Unified Policy for Position and Force Control in Legged Loco-Manipulation

In this talk

1. 2:45

   Why force control matters

   Bong describes trying to make a Unitree robot wipe a whiteboard. Position-only teleoperation either applied too little pressure to erase marks or too much for the robot to follow the path. Legged robots often lack force sensors, and adding them can be costly and difficult.

2. 6:58

   Estimating force through position offsets

   UniFP models the robot as a mass-spring-damper system and estimates external force from the difference between commanded and actual position. It adjusts the position target to produce compliant responses, force tracking, or gravity compensation. The same idea can adapt a walking robot’s desired velocity when someone pushes or stops it.

3. 13:32

   Training and real-world performance

   The team trains a whole-body control policy in simulation with sampled position and force commands, external forces, and robot state histories. Bong says estimated forces were within roughly a five-newton range, which may be adequate for heavy lifting but too imprecise for delicate tasks such as holding an egg. Adding force information to imitation learning improved success by 39.5% over a vision-only Diffusion Policy.

4. 28:00

   Applying the approach to collaboration

   For collaborative carrying, the team trains a whole-body controller, then a residual teacher that responds to an object’s movement and stability, and distills it into a student policy for deployment. In demonstrations, a person guides the robot by pulling it to move and pushing it to stop. Bong notes that autonomy remains limited and that command selection still needs work.

5. 32:25

   Limits and next steps

   Bong identifies force estimation accuracy, position precision, and choosing between force and position commands as open problems. He expects several useful compliance levels to matter soon, while precise force estimation may take three to five years. Better sensors and faster feedback will be needed for demanding interactions such as juggling or delicate manipulation.