Robot training method practises only the subtasks a policy keeps failing

single source· 1 articles · confidence: high · first seen 2026-09-17 20:00 UTC

What this means for you

Nothing to act on yet: this is a preprint with no released code or weights, and the results come from the authors' own robot runs. Worth watching if you fine-tune manipulation policies — freezing the base policy and training only the failing subtasks is directly transferable.

PARTS, a real-world reinforcement-learning method for robot manipulation, trains only the subtasks a pretrained policy already handles badly, instead of collecting fresh full-task demonstrations. A frozen pretrained policy supplies the nominal actions, while learned selectors and success verifiers add residual corrections and score each subtask locally — so training gets a reward signal even when whole-task runs rarely succeed. Complete-task success rises from 32% to 61% on bimanual YAM tasks and from 50% to 95% on single-arm Franka tasks, at tens of minutes of rollouts each. The authors report over 25% higher full-task success than prior real-world RL fine-tuning at equal rollout budget; no independent replication is given.

Key facts

  • ·PARTS is reported to raise complete-task success from 32% to 61% on bimanual YAM tasks and from 50% to 95% on single-arm Franka tasks. source
  • ·Average training cost was tens of minutes of real-world RL rollouts per task. source
  • ·The authors report more than 25% higher full-task success than existing real-world RL fine-tuning methods under the same robot-rollout budget, with less human involvement. source
  • ·The method keeps the pretrained policy frozen and trains residual corrections, activated by agent-generated selectors and success verifiers that supply local outcome rewards. source
  • ·The work was posted as arXiv preprint 2609.21788 on 17 September 2026. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire