GPT-6.2 Astra Strikes Again at Embodied Models: Success Rate on 16 Tasks Doubles Compared to GPT-6 Astra

Also achieves world's first pass on deformable object manipulation
Heng Yu, reporting from Waifeisi | QuantumBit | WeChat account QbitAI
GPT-6.2 Astra has once again raised the bar for embodied intelligence, delivering a fierce meteoric punch.
In the latest real-world tests by robotics infrastructure company XDOF, GPT-6.2 Astra's combined success rate in controlling a dual-arm robot to complete 16 manipulation tasks soared from 19.6% for GPT-6 Astra two months ago to 41.2%, more than doubling.
The two tasks of knot-tying and T-shirt unfolding, which no AI had previously passed, were also conquered this time.

After the results were announced, the official OpenAI Developers account reposted the news, and OpenAI co-founder and president Greg Brockman also spoke proudly:
GPT-6.2 Astra performs extremely well in robotics~~

To ensure evaluation rigor, XDOF ran six YAM workstations in parallel, repeating each task thirty times, completing 480 real tests per day, with manual annotation of sub-steps, grasp failures, and collisions, filtering out hardware bias and random disturbances.
Even with such stringent standards, XDOF stated in its latest tweet that the Ember-0 benchmark may soon be saturated by model capabilities.
Doubling Performance in Two Months—How Strong Is It Really?
To understand the significance of this breakthrough, one must first grasp just how difficult the test tasks are.
GPT-6.2 Astra participated in the Ember-0 test launched by XDOF.
This is a robotics evaluation benchmark oriented toward real physical environments, using a dual-arm robot to complete 16 manipulation tasks, each executed under a protocol of 30 repeated trials.
The tests cover capabilities including spatial reasoning, object grasping, dual-arm coordination, tool use, fine manipulation, and deformable object handling.
Two months ago, GPT-6 Astra's combined task success rate on the Ember-0 test was 19.6%.
By GPT-6.2 Astra, this number had soared to 41.2%, an increase of 21.6 percentage points, more than double the original—the 41.2% here is the equal-weighted average of success rates across the 16 shared manipulation tasks, covering object grasping, spatial reasoning, dual-arm coordination, tool use, and various other capabilities.
So, the progress did not come from a single-task breakthrough; GPT-6.2 Astra showed marked improvement across the entire test suite.

The most eye-catching were the two operations "tying a knot in a rope" and "unfolding a T-shirt," which GPT-6.2 Astra easily conquered.
Previously, Ember-0 had cumulatively conducted 143 valid T-shirt unfolding tests, and no model had successfully completed this task.
Beyond these two landmark breakthroughs, the model also showed marked progress on several other tasks covered by Ember-0.
- "Pulling a towel to retrieve a distant plush toy" and "tilting a tray so a ball rolls into a bowl," which require physical reasoning
- "Finding an occluded marker" and "following hidden instructions on a red sticky note," which test memory and perception
- "Moving a potato chip without crushing it," "picking up bread with tongs," and "carrying cereal onto a tray," which test fine force control
GPT-6.2 Astra demonstrated significant improvement in both checkpoint completion and final success rates across various sub-tasks.
The most notable improvement was in the task of picking up bread with tongs: GPT-6 Astra's success rate was only 10%, while GPT-6.2 Astra raised it to 90%.

Why Are Knot-Tying and T-Shirt Unfolding So Difficult?
To understand the significance of GPT-6.2 Astra's breakthrough, one must first grasp just how tricky the tasks of "knot-tying" and "T-shirt unfolding" are for robots.
In the field of robot learning, vision-language-action models (VLA) represented by Diffusion Policy, ACT, and RT-2, along with specialized policy models, typically rely on large amounts of human demonstration data to learn manipulation skills.
Humans demonstrate first, and the robot learns from this how to grasp, move, and place objects.
This approach has achieved good results with rigid objects and fixed procedures.
Objects like cups and boxes have relatively stable geometry, and grasp positions and motion trajectories are easier to predict; after training, models can often form relatively stable manipulation policies.
But, change the object, change the environment, or chain several actions into a more complex task, and the original policy may no longer work. The model may need to re-collect demonstration data and supplement training to adapt to new manipulation conditions.
Deformable objects raise the difficulty another level.
When unfolding a T-shirt or tying a knot in a rope, the object's shape changes continuously with each contact, grasp, and pull by the robotic arm, and the deformation process is hard to predict in advance. The robot needs to understand the task goal of "unfolding the clothes," and must also continuously judge the corners, wrinkles, and overall shape of the garment from visual feeds returned by wrist and overhead cameras, then decide how to act next.

This places high demands on manipulation precision and dual-arm coordination.
The slightest misstep in grasp position or action rhythm can cause the hem to slip, the rope to tangle, and previously completed operations to be undone.
In XDOF's evaluation system, the core capability tags corresponding to the tasks "knot-tying" and "T-shirt unfolding" cover deformation control, dual-arm coordination, dexterity, and long-horizon planning.
Under hardware conditions without tactile feedback, there is also an inference latency of more than ten to twenty-plus seconds between each action chunk output by the large model; the robotic arm cannot make real-time fine adjustments while waiting for new decisions, making the cost of every contact control and failure recovery extremely high.
Therefore, knot-tying and T-shirt unfolding test a comprehensive set of capabilities, including 3D spatial understanding, dual-arm coordination, fine contact control, dynamic state tracking based on history, and long-horizon task planning.

BTW, speed is also a challenge that second-system models have always faced. Previously, one criticism of such models was that the inference process takes too long, making it difficult to meet the real-time requirements of robot operation.
This time GPT-6.2 Astra also improved on speed.
During the model's release period, its average prediction latency dropped from 15 seconds for GPT-6 Astra to 10.1 seconds.
However, compared with other model architectures, this inference approach is still relatively inefficient. For robots that require continuous perception and rapid response, a wait of around 10 seconds is still not short.
But overall, embodied intelligence models often perform well on simple grasping and object placement; once they encounter complex contact, tool use, and deformable objects, success rates drop markedly.
The Ember-0 test results previously published by XDOF also showed similar problems.
For example, models can usually understand the task and plan seemingly reasonable actions, but when actually executed on the robotic arm, they tend to falter in grasping, contact, and continuous operation.
GPT-6.2 Astra's breakthrough this time precisely indicates that general models have begun to break through these difficulties.

From an industry perspective, the impact of general large models on embodied intelligence is also continuing along this trajectory.
From the NVIDIA research team testing GPT-6 Astra directly as a robot policy model, to Claude demonstrating complex manipulation capabilities in robot simulation environments, to Google DeepMind extending Gemini's multimodal understanding capabilities to robot control, general models are refreshing the capability boundaries of embodied intelligence round after round.
One More Thing
Jiafei Duan, a Presidential Young Professor at the National University of Singapore active in the robotics/embodied intelligence field, subsequently posted praising GPT-6.2 Astra's excellent performance in "invading" embodied intelligence.
In their real-robot comparison tests, GPT-6.2 Astra performed brilliantly, shooting to first place on the RoboColosseum SO-101 Elo ranking within a single day.
For the actual operation, please see the small screen:

From the initial model being helpless before deformable manipulation, to GPT-6.2 Astra doubling its success rate within two months and breaking through multiple zero-score tasks, the pace of evolution of frontier multimodal large models in the physical world is continuing to accelerate.
And at a pace that is somewhat overwhelming.
I declare that every day in November is a dizzying, collapse-into-a-chair day in the AI field (not really).
Reference links: [1]https://x.com/xdofai/status/2123822012066157011 [2]https://x.com/djiafei/status/2123925760281067567?s=46
- miHoYo published a VLA paper 2026-11-21
- Yao Shunyu and the Hunyuan team went on Reddit to answer netizens' questions, Hy small model already on the roadmap 2026-11-21
- Embodied intelligence going global, logistics leads the way 2026-11-09
- The Riemann Hypothesis proven?! With Claude's help, a 34-year-old mathematician releases a Lean proof 2026-11-21
Related Reading

Such an awesome world model, and it's open source!
Ant Lingbo's third release: LingBot-World
Article title: GPT-6.2 Astra Strikes Again at Embodied Models: Success Rate on 16 Tasks Doubles That of GPT-6 Astra
Body:

Autonomous Driving Veteran Yu Yinan's Robotics Startup: Nearly 100 Million in First Funding Round, Backed by Former Employers Horizon and Baidu

Robots Take On Extreme Outdoor Challenges Together in Hong Kong, Dogs Outperform Humans
Pre-race warning of a crash? Contestants: Doesn't exist

From Learning from Humans to Learning from Robot Companions: Tsinghua University, Together with Infinigence AI and AgiBot, Proposes a New Path for Scaling Embodied Intelligence Deployment
Let robots become each other's "teachers"

π0.7 Released, VLA Bets on Robotics' GPT-3 Moment
π0.7: A Controllable Model with Emergent Capabilities

Richard Liu Backs 3 Embodied Intelligence Companies in a Row! The JD–Meituan "War" Spreads Beyond Food Delivery
Betting on "mass-produced hardware + embodied intelligence brains"