AtomicRL lifts long-horizon VLA robot performance over PPO
Original title: AtomicRL: Atomic Reinforcement Learning for Long-Horizon Robotic Manipulation
Reported gains are 3.6–4.0% on LIBERO-LONG, up to 9.5% on CALVIN, and 21.3% on real robots.
Researchers introduce AtomicRL, a reinforcement-learning post-training method for vision-language-action policies on long-horizon manipulation. Semantic-Aware Policy Optimization limits credit contamination across atomic subtasks, and Bottleneck-Adaptive Target Exploration focuses learning on weak subtasks. A Real2Sim2Real pipeline is used to move simulation training onto physical robots.
- SAPO stops failed-subtask TD residuals from crossing task boundaries.
- BATE shifts sampling and exploration toward unmastered atomic subtasks.
- LIBERO-LONG improves 3.6% and 4.0% over PPO on two VLA architectures.
- CALVIN ABC-D full-task performance rises by up to 9.5% over PPO.
- Real-world tasks improve by an average of 21.3%.
From the original article
1 min read at arxiv.orgVision-language-action (VLA) models have shown promise in robotic manipulation, but imitation learning (IL) alone often struggles to master complex long-horizon skills. Reinforcement learning (RL) offers an effective avenue for further improving VLA policies.
By Likui Zhang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.