AI News

Anthropic and NVIDIA Release Robot-Related Models on the Same Day

East MoneyUpdated Nov 10 16:19

AI summary

On November 9 local time, Anthropic released Fable 5.2 and Mythos 5.2. Among them, Fable 5.2 achieved an average success rate of 64.9% on the RoboDojo dual-arm simulated robot benchmark test, compared with 27.6% for the previous generation Fable 5.1; during the same period, Claude Opus 5.5 scored 39.4% and GPT-6 Astra 29.8%. On the same day, NVIDIA launched the robot model GR00T 2 built on the Cosmos platform, as well as the robot learning framework Isaac Lab 3.0; GR00T 2 achieved an average success rate of 71% across 50 tasks with only 1 demonstration per task, rising to 95% with 5 demonstrations.

Why read this

AI score 71

Two companies released robot-related models on the same day; the benchmark comparisons and task success rate data given in the article can be used to assess the progress of general-purpose large models entering the embodied field.

Topics

Tags

Body · AI translation

Article body EM_StockImg_Start

EM_StockImg_End

On November 9 local time, Nvidia and Anthropic successively released new products, adding two more sparks to robotics.

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

Among them, as a general-purpose large model, the capabilities Fable 5.2 demonstrated in robot control drew industry attention. Test data released by Anthropic shows that on the RoboDojo dual-arm simulated 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 scored only 27.6% on this test; in other words, a "0.1-generation" update raised the success rate by 37.3 percentage points; and in the same-period comparison for Fable 5.2 this time, Claude Opus 5.5 scored 39.4%, while GPT-6 Astra scored 29.8%.

A higher success rate score also means the model is stronger at judging the environment from camera images and arranging actions.

Moreover, Fable 5.2 also possesses generalization capability. Anthropic pointed out that "Fable 5.2 shows greatly improved capabilities in controlling physical robots, even including tasks it was never explicitly trained on." In a demo video released by Anthropic, Fable 5.2 controlled a physical robotic arm to complete three tasks that had not been specifically trained for: drawing patterns, unscrewing bottle caps, and solving a Rubik's cube.

Anthropic said Fable 5.2's performance in robot control is "better than most AI models designed specifically for robot control."

This is also an important reason why Fable 5.2 sparked a huge response—previously, the mainstream view held that robots needed more vertical large models designed specifically for embodied intelligence to effectively solve spatial understanding, action arrangement, and control, whereas Fable 5.2, as a general-purpose large model, while excelling in complex programming, scientific research tasks, computer operation, interdisciplinary reasoning, and chart recognition, has also achieved a major leap in robot control.

On the other side, on the same day, November 9 local time, Nvidia announced the launch of its all-new next-generation robot model GR00T 2, and released the robot learning framework Isaac Lab 3.0 along with multiple robot training and evaluation technologies.

Nvidia said GR00T 2 is an open frontier world model built on the Cosmos platform. The model is trained using robot and human data, including 100,000 hours of open-source first-person human perspective data provided by Nvidia.

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

Previously, on November 5, Nvidia had just released four Cosmos 3 Nano open-weight models, covering multi-view scene generation, interactive simulation, and robot action prediction; Apple's research team, on November 4, released the HALO humanoid robot autonomous control system, which, by synthesizing training data in 1,000 procedurally generated home scenes, enables robots to autonomously perform tasks such as walking, obstacle avoidance, sitting down, and pushing boxes in real environments that have not been mapped.

It is worth mentioning that on November 4, Wang Xingxing, founder, chairman, and CEO of Unitree Robotics, said that 2027 has a very high probability of becoming the critical point for the robotics industry or robot model development, and that robot capabilities are expected to see substantial improvement next year; the "ChatGPT moment" of embodied intelligence is close at hand. And models are precisely the technical threshold for realizing this moment: AI models' understanding of physical contact and touch is still insufficient, and there is a misalignment between model inputs and outputs and the real world. He expects this problem to be resolved within the next one to three years.

A research report by Orient Securities on November 2 pointed out that the main reason the market was previously pessimistic about the embodied intelligence industry was the belief that embodied models are different from large models, and concern that slow model progress among embodied industry players would lead to limited application scenarios for embodied intelligence. However, as GPT-6 Astra has achieved good application practices in the embodied field, some investors have begun to gradually believe that general-purpose large models are likewise expected to capture market opportunities in the embodied field, and that the entry of large-model vendors will accelerate progress in embodied models, so expectations for industry commercialization are continuously rising.

The analyst further pointed out that large models have huge potential as the embodied strategy layer, and embodied body manufacturers stand to benefit the most at the margin. As general-purpose large-model vendors continue to explore the embodied intelligence field in the future, this is expected to drive accelerated application of embodied intelligence in logistics and industrial manufacturing, and vendors with vertical-scenario know-how will have relatively higher barriers. It is expected that more large-model vendors will accelerate their layout in the embodied field in the future, which is expected to drive breakthroughs in more application scenarios, thereby driving accelerated breakthroughs in the industry's commercialization process, and robot complete-machine and component manufacturers are expected to benefit.

(Source: Cailian Press) Body middle embedded advertisement

Source

East Money