GPT-6 Astra Completed 7 of 100 Robotic Tasks. Its Closest Competitor Completed Zero.

GPT-6 Astra completed 7 out of 100 robotic desk-manipulation tasks in dual-arm trials. Its closest competitor, Ai2’s MolmoAct2, completed zero. That gap — between completing tasks at a low rate and failing to complete any — is what researchers at Cornell and Google DeepMind are calling a “step change” in spatial reasoning.

What Happened

The StationeryBench benchmark tests models on real-world physical manipulation: a dual-arm robot interacts with objects on a desk — pens, notebooks, cups, typical surface clutter — across 200 trials. The benchmark is designed to evaluate whether a model can understand three-dimensional spatial relationships well enough to guide physical manipulation.

Results:

• GPT-6 Astra: 7 of 100 tasks completed fully; median progress score of 46 out of 100
• MolmoAct2 (Ai2): 0 of 100 tasks completed; median score of 12 out of 100

A separate benchmark, REMAP, showed Astra reaching accuracy near human-level performance in specific spatial reasoning scenarios, though performance varies across different task types.

The research team’s hypothesis: OpenAI trained Astra on large-scale 3D data including Blender scenes and similar synthetic spatial environments, which would explain its specific advantage in tasks requiring three-dimensional understanding.

Why It Matters

The absolute performance numbers are low — 7% task completion is not production-ready robotics. What matters is the distance between Astra and the field. Zero task completion vs. seven percent represents a qualitative gap: MolmoAct2 cannot complete the task at all, at any trial. Astra can complete some of them, some of the time. That’s a different category of performance, not a difference of degree.

This is the pattern that has historically preceded meaningful deployment: a model that can sometimes do something others cannot do at all is the early indicator of the capability that eventually becomes reliable.

What This Means for Builders

Robotics and physical AI are no longer purely hardware problems. The bottleneck in deploying useful robotic systems has consistently been the model’s ability to understand and reason about physical space. If Astra’s spatial reasoning advantage is real and durable, it means software capability has caught up with the hardware in ways that matter for the next generation of robotic applications.

3D and spatial applications have a new capability floor. Beyond robotics, spatial reasoning applies to AR/VR scene understanding, architectural and engineering design tools, and simulation environments. The same capability that guides a robot arm through desk clutter can power better spatial reasoning in non-physical applications — understanding floorplans, reasoning about physical systems, or working with 3D design files.

Agentic systems that interact with the physical world gain a new viable foundation. Agentic AI has primarily operated in digital environments: browsers, codebases, APIs, data pipelines. Physical agentic tasks — picking, placing, assembling — require spatial reasoning at a level that wasn’t reliably available before. The StationeryBench gap suggests that gap is narrowing for Astra-class models.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.