Event
RoboEvolve Releases v2, Extending to World-Action Models
1 reports · 1 sources
Harold Haodong Chen and other authors released v2 of the RoboEvolve framework paper, extending this planner-simulator co-evolution framework to the world-action model setting and adding new LIBERO low-level control experiments. The framework couples a VLM planner with a VGM simulator into a mutually reinforcing closed loop, bootstrapping from unlabeled seed images to alleviate manipulation data scarcity. The authors claim that only 500 unlabeled seed images suffice to surpass fully supervised baselines, with initial reinforcement learning data reduced by 50 times; on LIBERO, using environment feedback, the average task success rate of the cold-started LingBot-VA is raised from 78.7% to 98.7%. These are all results reported in the paper.
Report timeline
- RoboEvolve: A Planner-Simulator Co-Evolution Framework for Robotic Manipulation
RoboEvolve couples a VLM planner and a VGM simulator into a mutually reinforcing co-evolutionary loop, bootstrapped from unlabeled seed images to alleviate manipulation data scarcity. Experiments show it raises the base planner by 30 absolute points and improves simulator success by 48% on average, surpassing fully supervised baselines with only 500 unlabeled seed images and reducing initial reinforcement learning data by 50 times.
arXiv cs.RO ·