A robot arm must do more than briefly touch its target: precise manipulation also requires staying there. PolicyIK, a reinforcement-learning controller described in an arXiv preprint posted November 3, 2026, targets both behaviors. Researchers affiliated with New York University and NYU Abu Dhabi report improved reaching and holding in their benchmarks. Peer review has not been verified.
Inverse kinematics translates a requested tool position and orientation into joint motion. Multiple joint configurations can produce the same pose, while joint limits, gravity and controller behavior complicate execution. PolicyIK learns incremental changes to joint-position references through interaction. It retains the accumulated reference and observes its offset from measured joints, helping preserve the command needed to hold still. Coarse and fine error inputs, smaller actions near the target, and rewards for progress and time within tolerance support the final correction.
The main simulation benchmark used a seven-degree-of-freedom OpenArm in Isaac Sim, with 4,120 held-out start–goal pairs and three PPO training seeds. Each attempt lasted 5.3 seconds. Success required mean position and orientation errors over the final second to be no greater than 1 mm and 0.2°. PolicyIK achieved 93.49% ± 0.20 percentage points, with median terminal errors across all tasks of 0.25 mm and 0.04°. The reported variability is standard deviation across training seeds, not a confidence interval.
Tuned classical DLS, QP and MPC controllers achieved 50.41%, 56.21% and 56.33% success, respectively; an adapted UMI-RL baseline achieved 0% under the strict criterion. Goal corrections and retries raised PolicyIK’s result to 98.92% ± 0.23 percentage points, averaging 1.212 attempts per task. That figure includes additional attempts. Separately, without retries, 92.4% of tasks stayed within both thresholds at every sample of the final second, a stronger requirement than average-error success.
After direct hardware training with FlashSAC, PolicyIK reached tolerance on 54 of 60 physical OpenArm tasks within a 25-second budget and held it throughout the final five seconds on 43. Four classical comparators reached tolerance on 32–48 tasks and held it on 11–17. Hardware errors were calculated from joint encoders and forward kinematics; independent measurement of absolute physical tool accuracy was not established.
In ten simulated LIBERO-Object pick-and-place tasks driven by the π₀.₅ vision-language-action policy, PolicyIK achieved 95.7% success. Hand-tuned DLS achieved 96.7%, while default DLS and hand-tuned QP each achieved 90.8%. PolicyIK was retrained for the arm’s mounting configuration without pick-and-place training data. It therefore supported manipulation in this setting, but did not exceed the best comparator.
Simulation ablations support the design choices: removing reference-offset feedback reduced strict success by 9.2 percentage points; removing fine error inputs or adaptive action scaling cost roughly 4–5 points. These remain author-reported engineering results without supplied independent replication. Hardware and manipulation comparisons lack reported uncertainty or significance tests, and repeated runs do not represent independent robots. Reachable goals within a defined workspace also leave performance under unfamiliar obstacles, payload changes and unreachable commands unresolved.
Cover illustration generated with AI; conceptual, not research evidence.