跳到正文
返回论文列表
cs.RO提交于 已译

RoboEvolve:面向有限数据机器人操作的规划器—仿真器协同进化

RoboEvolve: Co-Evolving Planner-Simulator for Robotic Manipulation with Limited Data

Harold Haodong Chen · Sirui Chen · Yingjie Xu · Wenhang Ge · Ying-Cong Chen

中文摘要

机器人操作的可扩展性主要受任务对齐物理交互数据稀缺限制。视觉语言模型(VLM)与视频生成模型(VGM)虽可用于自主合成数据,却分别存在语义—空间错位和物理幻觉。RoboEvolve 将 VLM 规划器与 VGM 仿真器纳入相互促进的协同进化循环,并借助无标注种子图像启动自进化。其认知启发式双阶段机制包括:通过语义控制的多粒度奖励进行日间探索,以发现具物理依据的行为;通过挖掘“差一点失败”开展夜间巩固,稳定策略优化。自主渐进课程使系统从简单原子动作扩展至复杂任务。实验表明,RoboEvolve 将基础规划器提升 30 个绝对百分点,仿真器成功率平均提高 48%;仅用 500 个无标注种子便超过全监督基线,初始 RL 数据减少 50 倍;同时支持无灾难性遗忘的持续学习,并扩展至低层控制,借助环境反馈将冷启动 LingBot-VA 在 LIBERO 上的平均任务成功率从 78.7% 提升至 98.7%。

关键要点

  1. 01任务对齐物理交互数据稀缺,限制机器人操作扩展
  2. 02耦合 VLM 规划器与 VGM 仿真器,形成协同进化循环
  3. 03双阶段机制结合多粒度奖励与“差一点失败”挖掘
  4. 04仅用 500 个无标注种子,初始 RL 数据减少 50 倍
  5. 05借助环境反馈,将 LingBot-VA 冷启动成功率提升至 98.7%

解读

尚无解读。

原始英文摘要

arXiv:2605.13775v2 Announce Type: replace Abstract: The scalability of robotic manipulation is fundamentally bottlenecked by the scarcity of task-aligned physical interaction data. While vision-language models (VLMs) and video generation models (VGMs) hold promise for autonomous data synthesis, they suffer from semantic-spatial misalignment and physical hallucinations, respectively. To bridge this gap, we introduce RoboEvolve, a novel framework that couples a VLM planner and a VGM simulator into a mutually reinforcing co-evolutionary loop. Bootstrapped for self-evolution with unlabeled seed images, RoboEvolve leverages a cognitive-inspired dual-phase mechanism: (i) daytime exploration fosters physically grounded behavioral discovery through a semantic-controlled multi-granular reward, and (ii) nighttime consolidation mines "near-miss" failures to stabilize policy optimization. Guided by an autonomous progressive curriculum, the system naturally scales from simple atomic actions to complex tasks. Extensive experiments demonstrate that RoboEvolve (I) achieves superior effectiveness, elevating base planners by 30 absolute points and amplifying simulator success by 48% on average; (II) exhibits extreme data efficiency, surpassing fully supervised baselines with merely 500 unlabeled seeds--a 50X reduction in initial RL data; (III) demonstrates robust continual learning without catastrophic forgetting; and (IV) extends to low-level control, improving cold-started LingBot-VA from 78.7% to 98.7% average task success on LIBERO using environment feedback. Together, these results demonstrate the applicability of our co-evolutionary framework to both internally supervised generative learning and environment-grounded action learning.

同方向论文 · cs.RO

查看全部 →

反向发现 · 这篇论文相关的工具与提示词

去交叉发现 →

关联工具 · 1

Hugging Face Blog82

LeRobot v0.6.0 发布:引入世界模型策略、奖励模型 API 和新基准

LeRobot v0.6.0 发布,带来世界模型策略(VLA-JEPA、LingBot-VA、FastWAM)让机器人学会想象未来,以及一批新 VLA 模型(GR00T N1.7、MolmoAct2、EO-1、EVO1)。

推荐理由:原文给出了 world model policies、新 VLA 模型、reward models API 等具体能力变化,读者可以据此了解机器人学习工具链的最新进展。