Back to articles

GPT-6.2 Astra More Than Doubles Its Success Rate on Robot Tasks

3 min read

GPT-6.2 Astra has posted a sizeable improvement on Ember-0, a real-world manipulation benchmark run by robotics infrastructure company XDOF. According to the results shared by the company, the model reached an average success rate of 41.2% across 16 tasks performed by a dual-arm robot. GPT-6 Astra scored 19.6% on the same task set two months earlier. The increase is 21.6 percentage points and more than doubles the earlier result. It is a notable benchmark gain, but a 41.2% average also means that most attempts still do not end in full task success.

What the evaluation reports

  • A broad set of manipulation challenges. Ember-0 covers object grasping, spatial reasoning, two-arm coordination, tool use, dexterity and deformable-object handling. XDOF says each task was repeated 30 times across six YAM workstations. The evaluation also logged subtasks, grasp failures and collisions, an effort to separate model behavior from hardware issues and random variation.
  • Progress on flexible objects. GPT-6.2 Astra reportedly completed both knot-tying and T-shirt unfolding. XDOF says 143 valid T-shirt-unfolding trials had previously been conducted without a model completing the task. The largest task-specific improvement highlighted in the report was picking up bread with tongs: success rose from 10% for GPT-6 Astra to 90% for GPT-6.2 Astra. These examples suggest that the gains were not confined to one headline task, though the number of trials remains important when interpreting any individual score.
  • Lower latency, but not yet real-time. The report puts average prediction latency at 10.1 seconds, down from 15 seconds for the earlier model. That is a meaningful reduction, yet a delay of around ten seconds can still hinder robots that need to perceive changes and adjust continuously. The challenge is sharper when a system lacks tactile feedback and must wait for a new action chunk before correcting its movement.

Why are knots and clothing difficult? Unlike rigid objects such as cups or boxes, fabric and rope change shape as they are touched, pulled and repositioned. A robot must infer where the garment’s corners and folds are, or how a rope is arranged, then coordinate both arms and choose the next action. A poor grasp or an unsuitable sequence can undo earlier progress. These tasks therefore combine visual state tracking, precise contact, coordinated movement and planning over multiple steps.

What the result does—and does not—show

The scores add evidence that general-purpose multimodal models can make progress on physical manipulation, not just describe tasks or propose plausible actions. Gains across several task types are more informative than a single isolated success. XDOF’s separate report of a strong result in a RoboColosseum comparison is another claim from the source material, but it should not be treated as a substitute for broader, independently replicated testing.

Important limits remain. The published numbers come from one benchmark, and they do not show how well the system transfers to different robot hardware, unfamiliar environments or tasks outside the test set. XDOF itself suggests Ember-0 may approach saturation as models improve, making more demanding evaluations necessary. For deployment, the combination of a 41.2% average success rate and roughly ten-second prediction latency leaves a substantial gap between a promising benchmark result and dependable everyday robot operation.

Source: QbitAI

Comments

Checking sign-in status...

Loading comments...

Related articles

CCTest · Blog
ZeroVLN Navigates Without Navigation-Specific SFT of Its VLM Brain
Robotics & Physical AI
cctest.ai

ZeroVLN Navigates Without Navigation-Specific SFT of Its VLM Brain

ZeroVLN separates semantic movement decisions from continuous control: a vision-language model issues natural-language tool calls, and an action expert carries them out. The paper reports stronger results than the compared hierarchical methods on three benchmarks, without navigation-domain supervised fine-tuning of the VLM brain.

Read more