Anthropic and NVIDIA Unveil New Robot Control Models on November 9

“Science and Technology Innovation Board Daily” November 10 news On November 9 local time, NVIDIA and Anthropic successively released new products, adding two more fires to the robot field.

First, Anthropic, which was under scrutiny by the capital market on the eve of its IPO, released two major new models: The new generation flagship models Fable 5.2 and Mythos 5.2.

Among them, As a general large model, the capabilities demonstrated by Fable 5.2 in robot control have attracted industry attention. Test data released by Anthropic shows that in the RoboDojo dual-arm simulation robot benchmark, Fable 5.2 achieved an average success rate of 64.9%.

What level is 64.9%? The previous generation large model Fable 5.1 had a success rate of only 27.6% in this test; in other words, the “0.1 generation” update success rate increased by 37.3 percentage points; while in this Fable 5.2 period comparison, Claude Opus 5.5 scored 39.4%, and GPT-6 Astra was 29.8%.

A higher success rate score also means that the model's ability to judge the environment and arrange actions based on the camera footage is stronger.

and Fable 5.2 also possesses generalization capabilities; Anthropic pointed out, “Fable 5.2 demonstrates significantly improved capabilities in controlling physical robots, even including tasks that were never explicitly trained.” In the demonstration video released by Anthropic, Fable 5.2 controlled a robotic arm to complete three tasks without specialized training: drawing patterns, unscrewing bottle caps, and solving a Rubik's Cube.

Anthropic stated that Fable 5.2's performance in robot control is “superior to most AI models specifically designed for robot control.”

This is also an important reason why Fable 5.2 triggered a huge reaction—prior to this, the mainstream view was that robots need more vertical, large models specifically for embodied intelligence to effectively solve spatial understanding, action planning, and control, while Fable 5.2, as a general-purpose large model, achieves significant leaps in robot control while also demonstrating outstanding performance in multiple aspects such as complex programming, scientific research tasks, computer operations, cross-disciplinary reasoning, and chart recognition.

On the other side, on the same day, November 9 local time, NVIDIA announced the launch of the brand-new next-generation robot model GR00T 2, and released the robot learning framework Isaac Lab 3.0 and several robot training and evaluation technologies.

NVIDIA stated that GR00T 2 is an open-ended frontier world model built on the Cosmos platform. This model is trained using robot and human data, which includes 100,000 hours of open-source first-person human perspective data provided by NVIDIA.

One of the main features of GR00T 2 is its out-of-the-box task execution capability: out of 50 tasks, 40 can be completed directly without specialized demonstration training for specific tasks. Additionally, the model supports robots mastering skills by learning human operation demonstrations and can transfer the learned skills to robots of different forms without needing to re-perform task demonstrations for each robot form. According to NVIDIA's introduction, with only 1 demonstration provided for each task, GR00T 2 achieved an average success rate of 71% across 50 tasks; if 5 demonstrations are provided for each task, the average success rate increases to 95%.

Previously on November 5, NVIDIA just released 4 Cosmos 3 Nano open-weight models, covering multi-view scene generation, interactive simulation, and robot action prediction; the Apple research team released the HALO humanoid robot autonomous control system on November 4, achieving autonomous execution of tasks such as walking, obstacle avoidance, sitting, and pushing boxes in unmapped real environments by synthesizing training data in 1000 procedurally generated home scenes.

It is worth mentioning that on November 4, Wang Xingxing, founder, chairman, and CEO of Unitree, stated,2027 will very likely become the critical point for the robot industry or the development of robot models, and robot capabilities are expected to see a significant leap next year.; the “ChatGPT moment” of embodied intelligence is already within reach. And the model is precisely the technical threshold to achieve this moment: AI models' understanding of physical contact and touch is still insufficient, and there are deviations in the alignment between model inputs/outputs and the real world. He expects that this problem is expected to be solved within the next one to three years.

Orient Securities' research report on November 2 pointed out that, the main reason why the market was previously relatively pessimistic about the embodied AI industry was the belief that embodied models are distinct from large models, and concerns that slow model progress by embodied industry manufacturers would lead to limited application scenarios for embodied AI. However, as GPT-6Astra has already achieved good application practices in the embodied field, some investors have begun to gradually believe that general large models also have the potential to seize market opportunities in the embodied field, the entry of large model manufacturers will accelerate the progress of embodied models, and therefore expectations for industry commercialization are continuously rising.

Analysts further pointed out that large models have great potential as an embodied strategy layer, and embodied body manufacturers will benefit the most. As future general large model providers continue to explore the field of embodied intelligence, it is expected to drive the accelerated application of embodied intelligence in logistics and industrial manufacturing fields, and the barriers for manufacturers with know-how in vertical scenarios will be relatively higher. It is expected that in the future, even more large model providers will accelerate their layout in the embodied field, which is expected to push through more application scenarios, thereby accelerating the breakthrough of the industry's commercialization process, and robot complete machines and component manufacturers are expected to benefit.