GPT-6 Astra zero-shot performance on 100 BEHAVIOR tasks

Setup shared by every rollout

Videos show the head-camera view at its recorded resolution and frame rate; grid previews are 8-second time-lapses. The 100-task mean combines 99 official BEHAVIOR scores with a video-estimated Q ≈ 0.50 for the incomplete T08 rollout. Each episode includes the exact prompt delivered to the agent.