Robot-arm inverse kinematics moves to a learned closed loop: PolicyIK holds 1 mm on 93.5% of held-out tasks in one attempt
Synopsis
Robot-arm inverse kinematics is usually handled by a numerical solver plus a hand-tuned controller; PolicyIK replaces that stage with a goal-conditioned reinforcement-learning policy that updates joint references from pose feedback through a joint-position interface. On a simulated 7-DoF OpenArm, the trained PPO policy keeps the mean error over the final second within 1 mm on 93.5% of held-out tasks in a single attempt, and a deployment protocol raises success to 98.9%; trained directly on a physical arm, it reaches and holds more tasks than four classical IK controllers.
Interpretation
Robot-arm inverse kinematics can be handled without a numerical solver plus a hand-tuned controller, by a policy that learns from interaction and outputs joint-position references directly, with millimeter-level full-pose precision as the training target. Earlier learned reaching policies were measured on position goals or centimeter-level full-pose tracking, and when sub-millimeter accuracy was needed they were paired with a damped least-squares solver for the final correction. On a fixed-base 7-DoF OpenArm in Isaac Sim over 4,120 start-goal tasks disjoint from training, PolicyIK reaches 93.5% strict success in a single attempt with a median position error of 0.25 mm; the task-adapted UMI-RL baseline reaches 16.4% coarse success and no strict success in the same setting.
The policy accumulates its actions into the joint reference and observes the offset between that reference and the measured joints, so a zero action retains the offset the position controller needs at rest. If actions were applied relative to the measured joints, the policy would have to regenerate that offset with nonzero actions at every step, and the arm would sag whenever actions relax. Removing command-state feedback lowers strict success by 9.2 points; re-anchoring the reference at the measured joints keeps 95.9% coarse success but drops strict success from 84.3% to 30.8%.
Pose errors enter the observation at centimeter and sub-millimeter scales, the action step shrinks as the error falls, and the reward pays both for a new best error and for time spent inside the tolerance. A single action scale sets both transport speed and terminal resolution, so with a fixed scale small steps slowed transport and large steps lowered terminal precision. Removing the fine channels raises median position error from 0.25 to 0.36 mm, and replacing the error-dependent gain with a fixed gain raises median entry time from 1.73 to 2.28 s; without the dwell reward, 7.8% of tasks that enter the tolerance are not held through the final second, against 2.4% with it.
The same policy can be trained directly on a physical arm with an off-policy learner and can execute a vision-language-action policy's end-effector commands in simulated pick-and-place. Classical controllers require damping, task weights and gains tuned for each robot, and leave execution of joint motion to the robot's controller. On the physical OpenArm over 60 full-pose tasks with a 25 s budget, PolicyIK reaches 54 and holds 43 through the final 5 s, while the four classical controllers hold at most 17; on ten adapted LIBERO-Object tasks it executes the VLA commands with 95.7% success.
Perspective
The demonstrated scope is full-pose reaching and holding on the fixed-base 7-DoF OpenArm and xArm7 arms, with simulation evaluation on 4,120 start-goal tasks disjoint from training and physical evaluation on 60 tasks with a 25 s budget; the pick-and-place results come from ten adapted LIBERO-Object tasks in simulation.
The deployment protocol raises strict success from 93.5% to 98.9% but costs 0.21 extra attempts per task on average, and the 45 tasks that still fail have goal configurations closer to joint limits. In simulation the policy executes VLA commands with 95.7% success, comparable to hand-tuned DLS at 96.7%, with the gap concentrated on grasp targets near the mount.