A technical report by Jiayi Su, Yixin Zheng, Mi Yan, Li Yi, Zhizheng Zhang and He Wang examines whether GPT-6 Astra can improve robotic manipulation by reviewing and correcting actions proposed by π₀.₅.
On ten selected RoboDojo tasks, with five cases per task, the hybrid achieved 48% success and a mean score of 62.60. Direct GPT-6 Astra control achieved 26% success; its mean score of 37.81 uses the 48 episodes with available native scores. Both success rates use all 50 cases. The lead author’s research page also summarizes these results.
What the results mean
The hybrid corrected 14.4% of executed control steps. This supports combining reasoning with a learned controller on the evaluated tasks.
However, comparisons with other models use reweighted published references, rather than reruns with matching seeds. These results establish neither a benchmark-wide lead nor general reliability in real-world factories. This is a technical report, and its evaluation scope matters.
