In 2027, Not Letting General Models Control Robots Isn't Cutting-Edge AI

Embodied intelligence is no longer a closed-door game. If physical AI is a necessary path to the forefront, then OpenAI, Anthropic, and Google will all try to go through it. DeepSeek is laying out its plans, and it's not a secret.

(GTC2026 Berlin)

Recently, Otterman admitted that OpenAI will create a humanoid robot, with a very cool demo planned for next year. Yesterday, Huang Renxun announced at the Berlin GTC event that he is collaborating with Agile Robot to release a robot, which will be launched in the first quarter of next year, setting an example for the industry. The former focuses on applications, while the latter targets platforms, both hinting that 2027 will be a pivotal moment for embodied intelligence.

NVIDIA is very clear that OpenAI and Anthropic are demonstrating new possibilities. A few days ago, Jim Fan, NVIDIA's Robotics Director, revealed on social media platform X that on the day of the GPT-6 Astra release, the team switched the backend model of the framework to Astra, instantly pushing all benchmark scores to around 99%. Now Opus-5.5 is slightly stronger than Astra, while Sonnet-5.5 offers better cost-effectiveness.

Previously, leading AI companies were not without interest in robots. Physical AI was once seen as a necessary path to AGI, embodied intelligence is a typical scenario that integrates physical and AI, and the market size is extremely large.

But due to various reasons, the 'ChatGPT moment' for embodied intelligence has never arrived. About five to six years ago, OpenAI disbanded its own robot team and shifted focus to investing in more vertically specialized embodied intelligence companies. The market seems to believe that this is a vertical field that must be specifically tackled; the high valuations based on scarcity have a certain degree of rationality.

However, today, embodied intelligence is emerging from the expansion of general-purpose large models. Leading AI companies have already taken notice of this last year. Although their approaches are not entirely consistent, none of them now view embodied intelligence as a completely independent vertical model problem.

At the beginning of last year, Google DeepMind switched from its previous RT model series to the Gemini Robotics series, applying general multimodal capabilities separately to robot motion control and embodied reasoning. As Ted Xiao, a senior research scientist at the company, put it, this was the first contact between frontier large models and robots.

Dario Amodei thinks about the issue from the perspective of safety. Since last August, Anthropic has successively tried using Claude series models to control Unitree quadruped robots (Project Fetch) and DJI drones (Project Pilot), step by step evaluating when Claude could end-to-end control robots. These projects are led by the Frontier Red Team. In their view, embodied intelligence is an overflow of frontier AI capabilities, and does not require specialized training; what has happened in cybersecurity is also happening in embodied intelligence.

As for OpenAI, it rebuilt its humanoid robot laboratory in San Francisco at the beginning of last year, collecting a large amount of data; recently, it has started recruiting positions related to robotics, spanning software, algorithms, hardware, and engineering manufacturing.

The overflow of frontier AI capabilities is not difficult to understand. Semantic understanding and visual generalization capabilities help models understand embodied tasks and their environments; reasoning and planning capabilities allow models to break down multi-step tasks; tool invocation and programming capabilities enable models to perform operations through external interfaces. More importantly, frontier AI is becoming increasingly adept at adjusting strategies based on failure outcomes and gradually improving in subsequent attempts.

However, before the end of this summer, what frontier AI is still more proficient at is being a co-pilot for VLA models; the success rate of direct action control is still not impressive. In a common composite architecture, V, the VLA proposes specific actions, and frontier AI decides whether to accept them, and can also modify or replace actions. On the standard LIBERO-40 operational benchmark, this combination significantly improves task completion performance.

In early September, the release of GPT-6-Astra was a key turning point. OpenAI's CEO Greg Brockman called it a “generational leap,” even stating that “we are now entering the AGI era.” However, at the time, OpenAI's marketing focus was on “self-evolution,” and it was the embodied intelligence industry that ended up causing a stir. Soon after, the embodied intelligence community saw a surge in research targeting GPT-6-Astra.

Those startups that have raised significant funding, are still debating whether VLA (Visual Language Action Model) or WAM (World Action Model) is the future of embodied models, and have yet to find a path to scalable deployment, have found that general-purpose multimodal foundation models can already directly achieve part of the work of VLA models. In some sense, this will reshape the valuation logic of embodied model startups.

The expansion of frontier AI into the physical world is already showing signs. Anthropic's research shows that since the Opus-4.5 model at the end of last year, the performance of general models on embodied tasks has seen a significant leap; completing the same robot simulation benchmarks, GPT-6 Astra achieved significantly higher success rates than GPT-5.5. Research from institutions such as SenseTime shows that if the execution budget is substantially increased, allowing Astra to run for more steps, its success rate will show a trend of scaling with testing.

The NVIDIA Jim Fan team replicated these studies. They had GPT-6 Astra play three roles: directly controlling the robot (Direct Policy), continuously controlling through an agent (Agentic Policy), or supervising a dedicated model (Orchestrated Policy).

As a result, the success rate of Astra directly generating robot actions reached 66.3%; the success rate further increased to 74.8% when Codex provided continuous context and tool support. Under the same benchmarks, these figures exceeded other vertical embodied models. NVIDIA's research also found that under the agent mode of GPT-6 Astra, the difference in task success rates between ambiguous and standard instructions was not significant; this aligns well with everyday scenarios.

(Success rate and completion score on RoboLab-120. Solid colors represent success rates; light-colored extensions represent completion scores; line segments represent 95% confidence intervals)

Of course, many studies also point out that GPT-6 Astra still has significant limitations. Its inference latency is still not sufficient to meet the real-time control needs of rapidly changing scenarios; the high token cost also makes it difficult to compete with humans or dedicated models in actual deployment; in tasks such as high-precision operations and dual-arm coordination, its performance is still lacking; its task understanding often remains limited to user instructions, and overlooks changes in environmental state deviations during the process of achieving the target instruction.

But are these issues inherent boundaries that general frontier models cannot overcome, or are they just thresholds that can be crossed with more time, computing power, and iterations?

According to early-stage dollar fund Alphaist's founding partner Chen Zhe, as relayed, at the IROS 2026, a top-tier robotics conference held on September 30, the industry formed three distinctly different views on how Astra will impact embodied intelligence. About one-third of people believe that language models will eventually solve everything; the majority of pragmatic people believe that general models and embodied models will eventually have distinct roles; the remaining people believe that general models will never truly solve problems in the physical world.

Frontier AI may accelerate the development of the entire physical AI, expanding the market size. The value distribution in the embodied intelligence industry may also be rewritten. If the capabilities demonstrated by Astra can continue to improve along the trend of scaling, the application boundaries of embodied intelligence may be further expanded. Enterprises providing the underlying support, including computing infrastructure, embodied systems, and dexterous hands manufacturers, may all benefit from this. At the same time, integrating embodied systems with frontier AI to lower the development and deployment threshold for robots, as well as providing better embodied Harness frameworks, may also become new competitive directions.

At last week's Workbench, Wang Xingxing, who has been under controversy recently due to the company's stock price plunge, admitted that the biggest mistake he made in the past decades was not believing enough in AI. He was very excited about Astra, believing that with the release of GPT-7 next year, it might “directly solve the embodied intelligence problem.” At yesterday's GTC Berlin, Huang Renxun hoped that future robot technology could adopt standard configurations as much as possible, which is the reference humanoid robot Agile One SN provided by NVIDIA and Agile Robotics for the industry.

As for companies that have previously focused on embodied intelligence brains or adopted a vertically integrated approach, the entry of frontier AI means they must re-examine their differentiated advantages. This advantage may come from tactile and mechanical data that general models find difficult to obtain, or from industry-specific process knowledge and validation loops in real-world scenarios. They should talk with Harvey in the legal industry and OpenEvidence in the medical industry.

The valuation logic of domestic embodied intelligence companies will also face a re-evaluation. This summer, Tang Wenbin, founder of Yuanli Lingji, once stated outright that all the open-source embodied models published by domestic companies are not as useful as Pi-0.5. Now, these companies face even greater competitive pressure. Packaging traditional automation tasks that can already be achieved into embodied intelligence scenarios will not prove their scalable commercial capabilities.

Next year, domestic frontier AI open-source vendors will also face new challenges. So far, Alibaba's Rynn series, Byte's GR series, and ZhiPu's ZR-0 have mainly continued along the path of specialized embodied models. Will they attempt to directly integrate general models into robots? Meanwhile, DeepSeek, which is focused on AGI, has already participated in Unitree Technology's strategic placement and has invested in two underlying technology companies founded by former Tencent employees, seemingly with some ulterior motives.

--

Reference:

https://www.anthropic.com/research/project-fetch-robot-dog

https://www.anthropic.com/research/project-fetch-phase-two

https://www.anthropic.com/research/project-pilot

https://www.anthropic.com/research/claude-plays-robotics

https://research.nvidia.com/labs/srl/projects/astranomy

An Unexpected Robot Policy:Early Evaluations of GPT-6 Astra on RoboDojo and Beyond

Systematically Exploring the Capabilities of GPT-6 AstraasEmbodiedPolicies

GPT 6 Astra as an Embodied Policy