Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies

Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies. By Galbot Team and thirty-three others, from Galbot. Published on the twenty-ninth of September. Its code repository is worked out: about three hundred and fifty source files, with a readme and a licence.

GPT-6 Astra can generate numerical robot actions, which extends its role beyond high-level planning. The authors set out to test how far that goes, evaluating Astra as a general-purpose embodied policy across six domains. They look at three modes: direct control, cooperation with learned policies, and adaptation driven by feedback. Their conclusion is a gap. Astra is often good at deciding what should happen, and much less reliable at making it physically happen.

In gripper manipulation, Astra can correct task targets and set up contact conditions for a learned policy to execute afterwards. Hybrid control with the pi zero point five policy reaches forty-eight percent success on the evaluated RoboDojo subset. In dexterous manipulation, hybrid control succeeds in half of ten experience-guided DexJoCo trials, while direct in-hand control struggles to coordinate finger contacts. In mobile manipulation, hybrid control reaches thirty-eight point seven percent success on RoboCasa365.

Navigation is Astra's strongest domain in the authors' local comparisons: ninety-two percent success on RxR instruction following, and eighty-two percent on HM3D object search, though the searches take substantial detours. Locomotion is where it breaks down. Dense motion-reference generation is unreliable, and none of five sequential attempts on a single obstacle course reaches the goal, even though stability and forward progress improve. In humanoid loco-manipulation, with pretrained whole-body controllers, Astra beats the baseline methods on thirteen of the thirty HumanoidBench tasks.

Cost compounds the gap. Across fifty RoboDojo instances per condition, policy-assisted control consumes six hundred and twenty-four point eight million tokens. Direct control consumes one point one three two billion. A thirty-second locomotion run takes two hundred and fifty model calls, each averaging thirty-nine point eight six seconds, with the physics paused while the model thinks.

These are bounded results, and the abstract marks the bounds. Several numbers come from subsets: the evaluated part of RoboDojo, ten DexJoCo trials, a single obstacle course. The navigation lead is over the authors' own local comparisons, not a general ranking. The best manipulation numbers come from hybrid control, where a learned policy does part of the work. And the locomotion timing was measured with physics paused during every inference call, so the world waited for the model rather than the other way round.

Six domains, one recurring pattern. Where Astra decides, corrects targets, or follows instructions, it is useful. Where it must produce fine, continuous physical control on its own, it falters, and the time and tokens it spends per decision put a further limit on practical use.