Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies
A study finds GPT-6 Astra useful for robot task decisions but unreliable for direct physical control.
Researchers evaluate GPT-6 Astra as an embodied policy across gripper, dexterous, mobile, navigation, locomotion, and humanoid loco-manipulation tasks. Hybrid control with pi0.5 reaches 48% on a RoboDojo subset, 50% on ten DexJoCo trials, and 38.7% on evaluated RoboCasa365 scenes, while navigation hits 92% on RxR and 82% on HM3D object search. Direct locomotion fails all five obstacle-course attempts. A 30-second run needs 250 model calls averaging 39.86 seconds, with physics paused during inference.
- Hybrid control beats direct finger and locomotion control.
- Navigation reaches 92% RxR and 82% HM3D success.
- Astra exceeds baselines on 13 of 30 HumanoidBench tasks.
- A 30-second locomotion run needs 250 calls averaging 39.86 seconds.
From the original article
1 min read at huggingface.co
GPT-6 Astra exhibits a remarkable ability to generate numerical robot actions, extending its role beyond high-level planning. To assess Astra's capabilities as general-purpose embodied policies, we conduct comprehensive evaluations across six domains, examining direct control, cooperation with learned policies, and feedback-driven adaptation.
By Galbot Team, Xuchuan Chen, Xiaoqian Cheng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.