type
Post
status
Published
date
Nov 22, 2026
slug
summary
tags
category
技术分享
icon
password

GPT-6.2 Astra Doubles Embodied AI Success Rate: A Breakthrough in Deformable Object Manipulation

Recent independent testing reveals significant advancements in generalist multimodal models applied to robotics. GPT-6.2 Astra has demonstrated a doubling of task success rates on embodied benchmarks compared to its predecessor, GPT-6 Astra, marking a notable shift in how large-scale models interact with the physical world.

Performance Metrics and Benchmarking Methodology

The data comes from the robot infrastructure company XDOF, which conducted rigorous real-world tests on the "Ember-0" benchmark. This evaluation suite focuses on physical environments and utilizes dual-arm robots to complete 16 distinct manipulation tasks. To ensure statistical significance and filter out hardware inconsistencies, the testing protocol involved running six parallel YAM workstations. Each task was repeated 30 times, allowing for a capacity of 480 real-world trials per day. Human annotators verified sub-steps, identified grasp failures, and documented collisions to maintain data integrity.
The results show a substantial improvement over the two-month gap between model releases. The composite success rate for the 16 shared operation tasks, calculated as an equal-weight average, jumped from 19.6% for GPT-6 Astra to 41.2% for GPT-6.2 Astra. This 21.6 percentage point increase indicates that the improvement is broad-based rather than isolated to a single task. The model showed marked progress across various capabilities, including spatial reasoning, object grasping, dual-arm coordination, and tool usage.

Milestones in Flexible Object Interaction

Perhaps the most significant development is the model's ability to handle deformable objects, a domain where previous AI attempts have largely failed. Two specific tasks—tying a knot in a rope and unfolding a T-shirt—were previously unsolved by any AI model in the Ember-0 history. Prior to this update, there had been 143 valid attempts to unfold a T-shirt, with zero successes. GPT-6.2 Astra successfully completed both of these operations, representing a global first for this type of deformable manipulation.
In other tasks requiring precise force control and physical reasoning, GPT-6.2 Astra also outperformed its predecessor. For instance, in the task of picking up bread using tongs, the success rate soared from 10% with GPT-6 Astra to 90% with the newer version. Other tasks, such as retrieving a plush toy with a towel or tipping a tray to roll a ball into a bowl, also showed consistent improvements in checkpoint completion and final success rates.

Technical Challenges and Latency Considerations

The difficulty of tasks like knot-tying and garment unfolding lies in the unpredictable nature of deformable materials. Unlike rigid objects such as cups or boxes, which have stable geometric shapes, soft materials change form with every contact and pull. Robots must continuously interpret visual data from wrist and top-mounted cameras to judge wrinkles, corners, and overall shape while executing long-horizon plans. This requires sophisticated 3D spatial understanding, fine motor control, and the ability to recover from minor errors without tactile feedback.
However, speed remains a hurdle for second-system embodied models. While GPT-6.2 Astra improved its average prediction latency from 15 seconds to 10.1 seconds compared to GPT-6 Astra, this delay is still considerable for real-time manipulation. During these latency windows, the robotic arm cannot make real-time micro-adjustments, which increases the cost of contact control and failure recovery. Despite this, the model's ability to plan reasonable actions that are subsequently executed accurately by the mechanical arm suggests that general-purpose models are beginning to bridge the gap between high-level reasoning and low-level physical execution.

Industry Implications and Validation

The success of GPT-6.2 Astra has attracted attention from the broader AI community. Greg Brockman, co-founder and president of OpenAI, commented on the model's performance in robotics. Additionally, Jiafei Duan, a Presidential Young Professor at the National University of Singapore, highlighted the model's capabilities in real-robot comparisons. In their tests, GPT-6.2 Astra quickly achieved the number one ranking on the RoboColosseum SO-101 Elo leaderboard.
These results align with a broader industry trend where generalist multimodal models are encroaching on the domain of embodied intelligence. As seen with tests involving models from other organizations in simulated and real environments, the ability of large language and vision models to serve as policy generators is expanding the boundaries of what is possible in physical automation. The XDOF team noted that the Ember-0 benchmark may soon become saturated as model capabilities continue to scale, signaling an era of rapid iteration in physical AI.
原文出处:量子位,https://www.qbitai.com/2026/11/511779.html
AI正在加速人类文明进程开源Karpo Decision 31B:在300模型评测中登顶,以极速决策重构浏览器Agent
Loading...