The new RoboColiseum season introduces four tracks—dexterous hand manipulation, whole-body coordination, fine manipulation, and cross-embodiment transfer—comprising a total of 50 tasks that are closely aligned with real-world applications. Evaluation results have been released for 8 Agents and 3 policy models. Claude Fable 5.2 claimed first place overall with a score of 18.30, leading in both fine manipulation and cross-embodiment transfer. Notably, all participating Agents scored their lowest on the dexterous hand track, which has emerged as a common weakness for large models, and policy models collectively outperformed Agents on this track. The new season implements a triple anti-cheating mechanism, separates evaluations into Agent and policy model categories, and incorporates execution efficiency into scoring. Large-scale controlled experiments verify that the evaluation system's simulation and real-robot task scores achieve a correlation of 96.86%, indicating high credibility. The new season challenge is now open.
