anonymous-report-421/GPT-as-Policy

Evaluation code and a bilingual technical report by Jiayi Su, Yixin Zheng, Mi Yan, Li Yi, Zhizheng Zhang, and He Wang. Jiayi Su and Yixin Zheng contributed equally; Li Yi, Zhizheng Zhang, and He Wang are corresponding authors. We evaluate GPT 6 Astra Direct and a hybrid policy that reviews and optionally corrects π₀.₅ actions on ten RoboDojo tasks, with five aligned cases per task. The hybrid policy achieves a mean Score of 62.60 and 48% success, with GPT corrections on 14.4% of executed control steps. GPT 6 Astra Direct achieves 26% success and a mean Score of 37.81 over the 48 episodes with available native scores. Both success rates use all 50 selected cases. Official-model comparisons are reweighted public references, not same-seed reruns. This is a history-free source release. It contains no credentials, private annotations, raw model sessions, original operations logs, simulator assets or model checkpoints. The website repository…