AtomicRL: Atomic Reinforcement Learning for Long-Horizon Robotic Manipulation
arXiv cs.AI · 22 Oct 2026, 03:45 AEST
AtomicRL is a new reinforcement learning framework designed to enhance robotic manipulation skills. It addresses challenges in credit assignment and exploration, leading to significant performance improvements over existing methods.
Key facts
- Vision-language-action (VLA) models have shown promise in robotic manipulation, but imitation learning (IL) alone often struggles to master complex long-horizon skills.
- To address these challenges, we propose AtomicRL, an RL post-training framework that leverages atomic task structures for reliable credit assignment and adaptive exploration.
- On LIBERO-LONG, AtomicRL achieves absolute performance gains of 3.6% and 4.0% over PPO across two VLA architectures.
- Combined with the Real2Sim2Real pipeline, AtomicRL further improves real-world task performance by an average of 21.3%.
Read aloud
AtomicRL is a new reinforcement learning framework designed to enhance robotic manipulation skills. It addresses challenges in credit assignment and exploration, leading to significant performance improvements over existing methods. Vision-language-action VLA models have shown promise in robotic manipulation, but imitation learning IL alone often struggles to master complex long-horizon skills. To address these challenges, we propose AtomicRL, an RL post-training framework that leverages atomic task structures for reliable credit assignment and adaptive exploration. On LIBERO-LONG, AtomicRL achieves absolute performance gains of three point six percent and four point zero percent over PPO across two VLA architectures. Combined with the Real2Sim2Real pipeline, AtomicRL further improves real-world task performance by an average of twenty one point three percent.
Sources in this cluster
- AtomicRL: Atomic Reinforcement Learning for Long-Horizon Robotic Manipulation
arXiv cs.AI · 22 Oct 2026, 03:45 AEST