Event

EBench evaluation of GPT-6-Astra controlling robots

1 reports · 1 sources

Yifei Yao and other authors released a technical report evaluating GPT-6-Astra as a robot control agent on EBench, using the same observation protocol and post-training strategy, and comparing two configurations: in direct control, the agent directly generates waypoint commands; the harness combined with the strategy can also call post-trained VLA or WAM strategies to execute actions. The report states that under direct control, the agent's overall performance is comparable to the post-trained strategy, but because strategy invocation is suboptimal, it executes fixed commands without subgoals, and grasping is interrupted when the agent takes over again, performance remains below the per-task upper limit of each individual component.