Object-based Force-space Learning

Robot-agnostic force policies; sim→real.

Robot-agnostic policies in force-spaceCompositional primitives (prismatic/revolute)Sim→real transfer
Force-space policy
Description
We use force based formulation to learn object centric manipulation policies using reinforcement learning. After training in simulation, we use inverse kinematics to transfer these policies to a variety of robots including Spot, UR5, Panda and Kinova.
Related Publications
  • 2025FLEX: A Framework for Learning Robot-Agnostic Force-based Skills Involving Sustained Contact Object ManipulationIEEE ICRA
    arXivWebsite
People
  • Shivam Goel
  • Shijie Fang
  • Wenchang Gao
Shivam Goel