PolicyIK: NYU trains RL for inverse kinematics on a real arm, hitting 0.5mm median error

Lianhuiq · x · 2026-11-07

NYU researchers present PolicyIK, a goal-conditioned RL policy for high-precision inverse kinematics trained directly on a physical robot arm.

Method:

Updates joint-position references from pose feedback and observes the accumulated reference relative to measured joints, retaining the offset the position controller needs at rest.
Two-scale pose errors with error-dependent action gains resolve millimeter residuals while preserving large motions.
Progress and dwell rewards teach the policy to reach the tolerance and hold it.

Results:

In simulation, a single attempt keeps final-second mean errors within 1mm / 0.2° on 93.5% of held-out tasks; a deployment protocol raises success to 98.9%.
On a real arm: median errors of 0.50mm / 0.12°, reaching 54/60 tasks and holding 43; it beats all four classic IK controllers under the same time budgets and holds goal poses on 2.5× more tasks than the best classic controller.
Executes π0.5 VLA end-effector commands with 95.7% success in simulated pick-and-place.

More from Embodied