RoboEvolve posts 50x data efficiency

Harold Haodong Chen and four co-authors posted a paper titled "RoboEvolve: Co-Evolving Planner-Simulator for Robotic Manipulation with Limited Data" to arXiv on May 13, 2026. The authors reported that their system could match fully supervised baselines while using just 500 unlabeled seed images — a claim the paper frames as a 50x reduction in required training data. (arxiv.org) Harold H. Chen's group described RoboEvolve as a framework that couples a vision-language planner with a video-generation simulator in a co-evolutionary loop intended to generate curricula and stabilize policy learning. (arxiv.org) The paper's abstract summarizes the method as a daytime exploration phase for behavioral discovery and a nighttime consolidation phase for mining near-miss failures. ### Who are the authors and where did they publish the work? The paper lists Harold Haodong Chen, Sirui Chen, Yingjie Xu, Wenhang Ge and Ying-Cong Chen as authors on the arXiv submission. (arxiv.org) The preprint is indexed as arXiv:2605.13775 with a submission timestamp of May 13, 2026. ### What exactly do the authors mean by "50x data efficiency"? The paper states RoboEvolve "surpasses fully supervised baselines with merely 500 unlabeled seeds," which the authors present as roughly a 50x decrease compared with the data requirements of those baselines. (arxiv.org) The arXiv abstract quantifies gains as raising base planners by 30 absolute points and increasing simulator success by 48% on average. ### How does RoboEvolve work in practice? The authors describe RoboEvolve as a co-evolving planner-simulator system that operates from unlabeled images and uses an autonomous progressive curriculum to scale from atomic actions to complex tasks. (arxiv.org) The paper details a two-phase cycle — an exploration phase that produces physically grounded behaviors and a consolidation phase that stabilizes policy optimization by focusing on near misses. ### Which tasks and benchmarks did the team evaluate? The paper's abstract reports "extensive experiments" and summarizes average improvements across planner and simulator metrics; the arXiv listing and accompanying paper pages present those results as aggregate figures rather than naming every benchmark in the abstract itself. (arxiv.org) The preprint and linked paper text contain the experimental tables and benchmark names used by the authors. ### Did the authors release code, data, or training details? The authors and third-party paper listings indicate accompanying code and examples were placed on GitHub and related repositories tied to RoboEvolve accounts. (arxiv.org) A public backup repository and organization names matching the project appear on GitHub; the arXiv abstract and aggregated paper pages reference reproducibility materials alongside the preprint. ### Who has commented or validated the claims so far? Hugging Face and several paper-aggregation sites indexed the preprint and echoed the core claims immediately after the arXiv posting; independent peer review or community benchmarks have not been published in the public record linked to the preprint as of the arXiv posting. (arxiv.org) The authors label the submission as ongoing work. The arXiv preprint (arXiv:2605.13775) and the RoboEvolve-associated GitHub repositories are the primary public materials for replication and review; readers can access the paper on arXiv and the code repositories referenced by the authors for implementation details. (arxiv.org) The paper's arXiv DOI and the authors' GitHub repositories remain the next named milestones for public inspection and reproduction. (huggingface.co)