Force-aware / Compliant Policy Learning

Letting a learned manipulation policy sense contact force or output stiffness, so it maintains contact without pressing too hard.

This is an umbrella term for methods that let learned manipulation policies handle contact force. Common visuomotor policies (such as diffusion policy) only output a target position, executed by a rigid position controller, which tends to either press too hard or fail to stay in contact once touching something. Improvements take roughly two routes. One is ‘force-aware’: feeding 6-axis force/torque or tactile signals in as input, fused with images before outputting an action — for example, Shanghai Jiao Tong University's Cewu Lu group's FoAR (2024) weights the force signal based on a predicted contact state, and the same group's RDP (2025) pairs a low-frequency diffusion policy with a high-frequency tactile feedback loop, while ForceVLA (2025) adds a force-aware mixture-of-experts module inside a VLA. The other is ‘compliant’: having the policy also output parameters like stiffness, executed by an impedance or admittance compliant controller — for example, Stanford and TRI's Adaptive Compliance Policy, ACP (2024), outputs a reference pose, a virtual target pose, and stiffness. The difficulty is that most teleoperation systems have no force feedback, making it hard to get good force and stiffness labels from demonstration data.

ExampleACP's vase-wiping task: the policy uses images and force signals to give, in real time, a reference pose, a virtual target pose, and stiffness along the compliant direction, executed by a compliant controller so the robot keeps its wiping tool pressed against a curved vase surface without pressing too hard; the paper reports over 50% improvement on contact-rich tasks compared to position-only visuomotor policies.

Also called

Force-aware Policy, Compliance Policy

Related

Force Control · Impedance Control · Variable Impedance Control · Contact-rich Manipulation · Six-Axis Force/Torque Sensor · Force-aware Vision-Language-Action Model

Sources

Adaptive Compliance Policy: Learning Approximate Compliance for Diffusion Guided Control (arXiv:2410.09309)
FoAR: Force-Aware Reactive Policy for Contact-Rich Robotic Manipulation (arXiv:2411.15753)
Reactive Diffusion Policy: Slow-Fast Visual-Tactile Policy Learning for Contact-Rich Manipulation (arXiv:2503.02881)

As of

2025-05