AI News

Wang Xingxing moves up the embodied intelligence timetable, Liang Wenfeng accelerates robotics layout

HuxiuUpdated Nov 10 23:01

AI summary

At the Industrial Fair, Wang Xingxing shortened his most optimistic judgment of embodied intelligence's "ChatGPT moment" from two to three years to one to two years, and said that he had not believed in AI enough in the past. The article cites RoboDojo tests, saying that GPT-6 Astra achieved an average success rate of 22.48% across 42 robotic manipulation tasks, higher than GPT-5.5's 0.88% and DeepSeek-Flash's 1.92%, and that when paired with the specialized model π0.5 it rose from 26% to 48%. DeepSeek has already made a strategic investment in Unitree Robotics and is preparing to lead the investment in robotics AI company Winverse AI, with a post-investment valuation of about $400 million; Liang Wenfeng is also evaluating personally forming an embodied team.

Why read this

SelectedAI score 76

The article places Wang Xingxing's accelerated timetable alongside Liang Wenfeng's investment in robotics under the same set of model capability data, so readers can understand why the boundaries of embodied intelligence competition are shifting outward.

Topics

Tags

Body · AI translation

[[**Wang Xingxing has shortened his estimate of the embodied intelligence tipping point from two to three years to one to two years, and admits he underestimated AI; Liang Wenfeng is meeting intensively with embodied intelligence founders and building out robot investments. General-purpose models have begun directly controlling robot actions, and are changing the timetable and competitive landscape of embodied intelligence.**

**Key points**

**1. Wang Xingxing sharply moves up the embodied intelligence timetable**

Within two months, Wang Xingxing shortened his most optimistic estimate for embodied intelligence's "ChatGPT moment" from two to three years to one to two years, saying GPT-7 might "punch straight through the embodied intelligence problem," and admitting that "over the past decade or so, I haven't believed in AI enough."

**2. GPT-6 begins directly outputting robot actions**

GPT-6 Astra achieved an average success rate of 22.48% across 42 robotic manipulation tasks, far above GPT-5.5's 0.88% and DeepSeek-Flash's 1.92%; when paired with the specialized model π0.5, the success rate rose from 26% to 48%. But in dangerous-instruction tests, it entered actual attempts 97 times, refusing on safety grounds only 2 times.

**3. Liang Wenfeng accelerates his move toward embodied intelligence**

DeepSeek has strategically invested in Unitree Robotics and is preparing to lead the investment in robot AI company Winverse AI (post-money valuation of about $400 million); Liang Wenfeng is also evaluating forming an embodied team himself. Competitive pressure from GPT-6 has loosened his original placement of embodied intelligence at the very end of the AGI roadmap.

**4. The robotics industry's technology clock is starting to be set by outside model companies**

General-purpose models do not need to replace specialized robot models; as long as they enter the planning, checking, and error-correction stages, they can raise the system's ceiling. But robots still lack an "internet of hands"; real-machine data, safety constraints, and failure recovery remain the slowest and most expensive stretch, and the tipping point has not yet arrived. 2026-10-13 07:13 Wang Xingxing underestimated AI, Liang Wenfeng begins reassessing robots Embodied Business Outpost Quick Overview

This article is from the WeChat public account: Embodied Business Outpost, author: He An

In less than two months, Wang Xingxing moved the embodied intelligence timetable forward once.

On August 20, at the World Robot Conference, he judged that embodied intelligence's "ChatGPT moment" would come in two to three years at the fastest, or five to ten years at the slowest. But by the October 12 Industrial Fair, the most optimistic timeline had become one to two years. Wang Xingxing even said that when GPT-7 comes out next year, it might "punch straight through the embodied intelligence problem."

Wang Xingxing has worked on robots for more than a decade and knows how many rounds of engineering debugging separate a machine from completing a demo to entering the field.

Robots are not like ChatBot products, where a wrong answer can simply be regenerated; if a robotic arm reaches out incorrectly once, it may knock over an object that was not the target, or damage equipment.

Phoenix Technology, citing multiple sources, reported that Liang Wenfeng has recently been meeting intensively with embodied intelligence founders to discuss robots, large models, and interaction with the real world. DeepSeek is considering moving further into embodied intelligence, possibly by forming a team itself or by entering through investment.

In a previously circulated transcript of an investor exchange recording, however, Liang Wenfeng gave a different order for the AGI roadmap: language models, chain of thought, agents, continual learning, self-iteration, and only then embodied intelligence.

The two paths are approaching from opposite directions. What is really worth asking is not who has caught up with another trend, but which shared premise has changed.

In embodied intelligence, the common approach had been to let a large model understand human instructions and break down tasks, then hand goals such as "pick up the cup" or "move to the sink" to a specially trained robot model.

The large model is like a dispatcher standing inside the brain, knowing what needs to be done; exactly how to raise a hand, move, and grasp is still decided by the robot's cerebellum.

GPT-6 Astra crossed a boundary.

It began directly giving the robotic arm's next position, orientation, and gripper state, after which the low-level control system converts these targets into joint motion.

The RoboDojo test published in September 2026 provided a set of easily overlooked comparisons.

The researchers did not fine-tune GPT-6 Astra for specific tasks; instead, they had it act directly as the action policy across 42 robotic manipulation tasks. After 2,100 tests, Astra's average success rate was 22.48%, outperforming the 40 public robot policies in the evaluation.

22.48% is far from practical.

But viewed another way, under the same post-processing scheme, GPT-5.5 and DeepSeek-Flash had average success rates of only 0.88% and 1.92%, respectively; among them, DeepSeek-Flash was tested fewer times per task.

The difference between frontier models is not getting two or three more tasks right, but that the route of "letting a general-purpose model directly output robot actions" has begun to genuinely work across a considerable share of tasks.

Capability distribution remains very uneven.

Astra performs better on tasks involving semantic understanding and low precision requirements; once it enters precision manipulation, dynamic control, or complex two-arm coordination, the success rate drops rapidly.

It can understand "build a tower," yet when reaching for the next block it may knock over the part already built.

These results do not prove that general-purpose models can replace robot models. What they change is another question:

Must a robot learn, bit by bit from its own data, the entire ability to go from understanding a task to deciding on actions?

Another test of 50 cases produced a more realistic combination: the specialized action model π0.5 was responsible for proposing actions, and GPT-6 Astra was responsible for checking and revising them, raising the success rate from 26% when Astra controlled alone to 48%. Astra actually revised only 14.4% of the actions.

A general-purpose model does not have to replace a robot model. As long as it enters the planning, checking, and error-correction stages, it may change the ceiling of the original system.

This capability even moves faster than safety mechanisms.

RoboHarm connected GPT-6 Astra to a real robotic arm and had it face dangerous instructions such as putting a power bank into water, inserting a metal screwdriver into a toaster, and mixing bleach with ammonia.

In 100 tests, Astra entered an actual attempt 97 times, completed the target action 60 times, and explicitly refused for safety reasons only 2 times.

In the past, a robot failing to complete a dangerous task may simply have been a matter of insufficient capability. As models become stronger, the original "failure" will not automatically become safety; instead, it may become a more reliable dangerous action.

GPT-6 has not solved robots' precision, real-time control, and safety, but it has already pushed the industry's problems one step forward.

Previously, people were still asking whether general-purpose models could control robots; now what needs to be discussed is which actions it is suitable to take over, how it should cooperate with specialized models, and who is responsible for making the robot stop when it should not execute.

Wang Xingxing believed early on that the main problem limiting the spread of robots is not hardware, but software and intelligence.

In the past, robot companies needed to collect data and train models for each task, then adapt them to the body, gripper, and workstation. Changes in lighting, part offset, or a single failed grasp could all lead to a new round of debugging. What a robot could do depended mainly on how much data the company itself had accumulated and how many actions it had solved.

After general-purpose models are connected, robots still have to learn how this body exerts force and how to maintain balance, but they no longer have to start learning from "what is a cup and how many steps should the task be divided into." The semantic understanding and task-planning capabilities it acquires from text, images, and video can be partially reused by robots.

At the China International Industry Fair, Wang Xingxing revealed that Unitree had already put GPT-6 on different robots to execute tasks, with results that were "very good."

More importantly, GPT-6's capabilities come from outside the robotics industry.

Unitree can continuously collect data and train models, but it cannot decide when the next generation of GPT will be released.

When external general-purpose models make a capability jump every so often, robot companies' R&D timelines will also be rewritten along with them.

This explains why, in less than two months, Wang Xingxing shortened his most optimistic judgment of the critical point from two or three years to one or two years.

In the past, the robotics industry had to fill in intelligence on its own; now, general-purpose models may cross part of the distance for it from the outside.

Robot companies still have to do the hardest and slowest work, but the technology clock of the robotics industry has already begun to be turned by model companies outside the industry.

3.

What Liang Wenfeng is looking for is not just a body

This also raises a question: since general-purpose models have already begun to affect robots' action capabilities, can Liang Wenfeng still leave embodied intelligence until the end of the AGI roadmap?

GPT-6 has loosened this sequence.

The body does not necessarily have to wait until AGI is complete before becoming the vehicle for models to enter the real world. When a general-purpose model can watch images, call control interfaces, and revise actions based on feedback, the robot itself also begins to become an environment for the model to continue learning.

Text can tell a model that "cups break easily," but how much force a grasp actually used, how the cup slid upon contact, and how to recover after a failed action can only receive more complete feedback through physical interaction.

Language models are good at predicting what content should appear next, while robots require the model to bear real consequences for an action.

This is especially important for the continual learning that DeepSeek cares about.

Therefore, Liang Wenfeng's approach to robotics does not necessarily mean that DeepSeek wants to become another Unitree.

For a model company, the value of robots is not just a new product that can be sold. It may be an entry point for general-purpose models to continue expanding capabilities, obtain real-world feedback, and test continual learning.

Looking again at DeepSeek's two moves at this point is very interesting.

One end is Unitree.

DeepSeek has already entered this robot company as a strategic investor, and the memorandum of cooperation signed by the two sides covers artificial general intelligence, high-performance general-purpose robots, and AI large models.

Unitree can provide the body, control system, real-world scenarios, and continuously generated real-machine feedback.

The other end is Winverse AI.

Bloomberg reported that DeepSeek is preparing to lead an investment in this robotics AI company, which was founded only a few months ago. The financing scale is tens of millions of dollars, with a post-investment valuation of about $400 million.

Founder Hu Han was previously Tencent's chief scientist, responsible for the vision-language model and embodied foundation model teams.

Judging from the choice of targets, what DeepSeek is approaching is not joints and complete machines, but the layer where visual understanding, foundation models, and robot actions begin to converge.

One end is the body, the site, and real-machine data; the other end is the model, the "brain," and action decisions. DeepSeek will not casually cast a net on a hot track; it looks more like it is positioning itself in advance at both ends of the embodied intelligence loop.

Add to that the report that Liang Wenfeng is evaluating personally forming an embodied team, and embodied intelligence's place in DeepSeek's roadmap has at least shifted from a distant endpoint to a direction that requires early positioning and continuous observation.

Winverse's financing has not yet been finalized, and DeepSeek has not officially announced the establishment of a robot product line, but capital moves have already run ahead of verbal statements.

The report mentions that competitive pressure brought by GPT-6 may be the catalyst.

The deeper change is this: if other frontier models have already begun to enter action decision-making, the data, interfaces, and evaluation standards of the physical world may be established by others first.

Waiting until all digital intelligence problems are solved before entering may not still be the easiest path.

Wang Xingxing hopes to use general models to shorten the distance for robots to become intelligent; Liang Wenfeng may need robots to give general models feedback that the digital world does not have.

The approach from both sides is not just "the brain meets the body."

4.

Why embodied intelligence may become hot again

Over the past two years, embodied intelligence has already gone through a wave of enthusiasm driven by financing, prototypes, and mass-production plans. But robots have become increasingly able to walk, run, and perform, yet after entering factories and homes they still lack generalization ability: the same action with a different object, position, or scene may require retraining. Hardware is advancing quickly, but intelligence has not kept up at the same speed.

The imagination brought by GPT-6 and DeepSeek is different from the previous round, and this is not an isolated echo between two Chinese entrepreneurs.

Google DeepMind and Nvidia are both pushing general models into robot control systems. More direct evidence comes from Cortex AI Research: it connected models such as GPT-6 Astra to the same set of dual-arm robotic arms. Astra can not only put a sandwich into a microwave, but also discover that the microwave door is not properly closed and that a pipette tip has not been attached, and proactively correct them.

This is not a performance designed for a specific robot, but an external team putting general models into unfamiliar bodies and unfamiliar tasks for comparison. Model companies, computing platforms, and robotics researchers are crowding into the same question: how does general intelligence acquire the ability to act?

Capital has also begun to price a "general robot brain" separately.

In January 2026, Skild AI completed a $1.4 billion financing round at a valuation of more than $14 billion; previously, Physical Intelligence raised $600 million at a valuation of $5.6 billion. Neither centers on manufacturing a particular robot; instead, they try to train foundation models that can transfer to different bodies and tasks.

These valuations may not yet be fully supported by commercial revenue, but they clearly reflect what capital is betting on: the pricing focus of capital for the robotics industry has begun to shift from complete machines, joints, and mass-production capability toward models, data, and cross-embodiment transfer capability.

Embodied intelligence has thus gained an engine from outside the industry that is still upgrading at high speed.

The release of a new general model may directly change the baseline for task understanding and action decision-making, forcing robotics companies to retest their original architectures.

Wang Xingxing moving his timeline earlier is a signal of this compression.

The competitive boundary will also shift accordingly.

Embodied model companies originally mainly compared data scale, cross-embodiment capability, and task success rates with peers. Now they also have to answer:

Once GPT, Gemini, or DeepSeek have similar capabilities, what irreplaceable value does their own model still provide?

The answer may be real-machine data, precision manipulation, low-latency control, failure recovery, and safety, or it may be long-term operating experience in certain scenarios, but it can no longer be just "we also have a large model."

Bodies, joints, and motion control will not lose value because of this.

The stronger the action intent given by a general model, the more important the work of converting it into stable, safe, and continuous physical actions becomes.

Robotics companies may thus diverge in position: companies that only provide a body are more easily compared; companies that can simultaneously connect to frontier models, accumulate real-machine feedback, and handle on-site delivery grasp the last segment that general models cannot complete alone.

5.

The timetable has changed, but the critical point has not yet arrived

The average success rate of 22.48% also means that, under RoboDojo's testing criteria, most tasks are still not completed.

The model's performance on simple placement tasks cannot be directly converted into qualified output for eight consecutive hours in a factory. Even less does it mean robots can enter homes.

The dangerous-action tests also show that capability improvement may even run ahead of safety constraints.

General models have also not given robots a free "action corpus."

Language models can obtain massive amounts of text from the internet, but robots do not have a ready-made "internet of hands."

Contact force, friction changes, object slipping, and failure recovery still have to be produced bit by bit from teleoperation, simulation, real-machine training, and on-site operation, and these data are often bound to specific bodies, sensors, and environments.

This determines that GPT-6 can rewrite the timetable for embodied intelligence, but for now cannot erase the slowest and most expensive part of embodied intelligence.

Therefore, Wang Xingxing's "biggest mistake" and Liang Wenfeng's "accelerating approach" are both insufficient to prove that the ChatGPT moment of embodied intelligence has arrived.

What they change together is the probability of this happening, and the way people estimate the time.

To judge whether this round of enthusiasm will turn into real progress, three results can be examined:

First, how much capability can be retained when the same model is switched to a different body and a different task.

Second, how qualified output, task time, and the number of manual takeovers change when robots work continuously.

Third, after model capabilities improve, whether the system can simultaneously establish reliable rejection of dangerous actions and boundaries of responsibility.

If these three items do not improve, new models can only produce more spectacular demonstrations.

If they improve simultaneously, the way the robotics industry progresses will truly change.

Now, the question is gradually becoming: how many more obstacles will the next generation of GPT and DeepSeek cross for robots?

Wang Xingxing, starting from the physical world, has re-estimated the speed at which AI solves robotics problems; Liang Wenfeng, starting from the digital world, has re-evaluated the position of robots in the AGI roadmap.

Neither has proved that the critical point has arrived, yet both have pushed the timetable forward by one step.

Embodied intelligence has not yet crossed the critical point, but the people who decide when it crosses may no longer be only within the robotics industry. AI industry signal channel: Frontier Technology. This content is published with the author's authorization; the views represent only the author and do not represent Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. People who are changing and want to change the world are all in the Huxiu APP [ ]

Source

Huxiu