AtomicRL: Atomic Reinforcement Learning for Long-Horizon Robotic Manipulation
AtomicRL targets weak subtasks in robot training, reporting an average 21.3% improvement in real-world task performance.
The framework post-trains vision-language-action models by using individual subtasks to guide reinforcement learning. Its policy optimization limits failed subtasks’ influence across task boundaries, while adaptive exploration focuses training on skills the robot has yet to master. Against PPO, the authors report absolute gains of 3.6 and 4.0 percentage points on LIBERO-LONG across two architectures, and up to 9.5 percentage points on CALVIN ABC-D. The real-world results combine AtomicRL with a Real2Sim2Real pipeline for transferring simulation training to physical robots.
ArXiv · AI/CL/LG's note
The framework post-trains vision-language-action models by using individual subtasks to guide reinforcement learning. Its policy optimization limits failed subtasks’ influence across task boundaries, while adaptive exploration focuses training on skills the robot has yet to master. Against PPO, the authors report absolute gains of 3.6 and 4.0 percentage points on LIBERO-LONG across two architectures, and up to 9.5 percentage points on CALVIN ABC-D. The real-world results combine AtomicRL with a Real2Sim2Real pipeline for transferring simulation training to physical robots.
ArXiv · AI/CL/LG's note
score 5