RoboColosseum：大规模真实环境下的物理智能评估框架
原标题：Paper page - RoboColosseum: Evaluating Physical Intelligence at Scale In-The-Wild
AI 摘要
研究团队提出 RoboColosseum，一个分布式机器人策略评估框架，通过统一策略服务与标准化协议，在 124 个站点、多种真实环境中对 5 个视觉-语言-动作（VLA）策略和 3 个前沿 AI 智能体进行评测，覆盖 SO-101、Franka、Bimanual-YAM 和 G1 Humanoid 四种本体。结果显示，前沿智能体在零样本条件下可与专用 VLA 策略匹敌，但所有系统在开放世界任务上都难以可靠完成；微调后 GR00T N1.7 在 SO-101 上领先，π0.5 在 Franka 和 Bimanual-YAM 上领先，且不同本体任务套件间的策略排名一致性很弱（Kendall W=0.24）。该工作揭示了机器人策略跨本体泛化与适应能力的评估难点，并提出主动匹配可减少物理试验次数。
正文节选
RoboColosseum: Evaluating Physical Intelligence at Scale In-The-Wild Abstract RoboColosseum evaluates general robot policies across diverse embodiments and environments, revealing weak cross-embodiment agreement, limited open-world reliability, and substantial differences between zero-shot and fine-tuned performance. General-purpose robot policies must work across robots, tasks, and environments, yet the variation that makes physical trials informative also makes them difficult to reproduce. We