Deep Dive — Robots are learning what their actions do, not what to do
Between May and August, at least ten Chinese embodied-AI companies shipped a world model. That is a land rush in four months, and it arrived just as the World Robot Conference in Beijing turned into a showcase for the same idea. When a field this young converges this fast, the useful question is not who is winning — it is whether everyone is building the same thing. They are not, and the split tells you more about robotics in 2026 than any leaderboard does.
The clearest statement of the split came this week from Guangxiang Technology (光象科技), a Tsinghua-incubated company that builds robots for car factories. Its position, in one sentence: a vision-language-action model learns what to do, a conventional world model learns what will happen next, and a robot working on a real production line needs a third thing — what will happen next because of what I did. The company calls the first-generation implementation of that idea ActEffect, and the contrarian part is what happens to it at the end of training: it gets fired.
Three questions that sound like one
The vocabulary collapse is real. "World model" now covers video generators that produce plausible fututes, latent dynamics models that compress a scene into a state vector, and joint-embedding architectures in the JEPA lineage that predict representations rather than pixels. A recent tutorial paper on the taxonomy notes that these differ in representation, predictive capability, and interaction mechanism — which is a polite way of saying two labs can use the same phrase and describe incompatible objects. Ten companies announcing "our world model" in one summer is partly a technical convergence and partly a labeling one.
The distinction Guangxiang is drawing sits underneath all of that. A conveyor belt moves on its own. The lights change. A part shifts a few millimetres from vibration. Meanwhile the gripper is actively changing things by pushing, grasping, lifting, and placing. Mix those together and a model can predict the next frame accurately while having no idea which changes came from the environment and which came from itself — which means it cannot answer the only question a controller actually asks: if I do something different, does it come out better?
Design follows from that. For states you can write down — joint angles, velocities, object poses, contact forces — the company reaches for existing physics knowledge rather than learning from scratch. For states you cannot write down, the ones buried in images, video, and point clouds, a neural network does the work. Ordering matters: close the gripper a few millimetres early, move a little faster, change the contact angle, and the downstream state has to change accordingly, so the model is trained to stay sensitive to intervention. Symmetry, conservation, and structure-preserving constraints get baked in so long rollouts don't drift into physically impossible scenes. Objects don't teleport; contact doesn't vanish.
Grade the consequences, then fire the grader
ActEffect's training loop is unusually concrete. The policy proposes three complete action candidates from the current visual state — complete matters, because diffusion- and flow-based policies pass through noisy intermediate states that can't be handed to a world model as executable actions. The controlled world model then predicts what the scene looks like after each of the three. The training data already contains the real future observation, so the three predictions get compared against it and ranked: the refined proposal must land closer to reality than the coarse one, and the coarse one closer than the plain feedforward pass. A ranking loss with gradient truncation keeps the model from gaming the comparison by making bad answers worse.
Then the interesting engineering decision. Once training finishes, the controlled world model and the future-observation branch are both removed from the deployment stack. Only the action head remains, doing a coarse proposal and a refinement in a single forward pass. The robot never expands futures at runtime and never searches over candidate actions. All of the "thinking about what this action will do" happened during training, and what survives is distilled into the policy weights.
The results, on standard simulation benchmarks: 98.8% average success on LIBERO, the four-suite, forty-task manipulation set; 80.3% on LIBERO-PLUS, which perturbs camera viewpoint, robot initial state, language phrasing, lighting, background, sensor noise, and object layout; and 67.5% on RoboCasa-GR1, a 29-dimensional action space across twenty-four dexterous tasks tested fifty times each for 1,200 total trials. The comparison points the company publishes are large — 80.3% against 51.5% for Fast-WAM on the perturbed set, and a 10.8-point margin over the same model on the humanoid-torso suite.
Read the ablations, not the headline
The honest reading of those numbers is in the ablation table, and it is less flattering than the headline. Strip out the consequence feedback and LIBERO falls from 98.8% to 97.0%. Swap the DINOv3 feature space for a vision-language representation and you get 97.3%. Remove the ranking loss entirely and you land at 98.1%. Those are real effects and they point the right direction, but a 1.8-point delta on a benchmark where the baseline is already at 97 is not the same claim as "world models unlock robot manipulation."
The 80.3% on LIBERO-PLUS is where the method earns its keep. A benchmark that moves the camera, the lighting, and the phrasing is measuring something closer to the failure mode that actually stops deployments — the same part in a slightly different place. That is exactly the problem Unitree's founder described when he put a number on the field last week, arguing that robotics takes off when a machine handles roughly 80% of tasks in roughly 80% of unfamiliar places; we covered that in Unitree's founder puts a number on robotics: two '80%' thresholds. An 80% score on a deliberately perturbed benchmark is not Wang Xingxing's 80%-in-80%-of-rooms, but it is the same shape of test, and it is the number worth watching rather than the clean-suite score.
Why firing the model matters more than building it
The deployment exit is the part that generalizes. In a lab you are graded on task success. On a line, every extra model pass becomes latency, power, hardware, and maintenance — multiplied across dozens of stations and hundreds of units. A robot that must imagine three futures before each grasp is a robot with a longer inference chain, a bigger GPU bill, and more things to fail at 3 a.m. Squeezing the consequence signal into the weights during training and shipping a single-forward-pass policy is a bet that the expensive part of understanding physics is a one-time cost.
It also reframes what a world model is for. The dominant framing is the onboard sandbox: the robot simulates, then acts. ActEffect treats the world model as a training-time critic instead — closer to how a human apprentice internalizes consequences and then stops consciously simulating them. If that works, the world model is not a runtime component competing for the compute budget. It is a way to extract more supervision from data you already collected.
That is a commercially legible argument for a company whose customers are carmakers. Guangxiang was founded in April 2025 out of Tsinghua's School of Vehicle and Mobility and its AI institute; CEO Zhang Tao came from autonomous-driving perception and localization, including a stint running the spatial perception engine at Alibaba's Amap. The company shipped an industrial robot, Phi-Bot X1, in June, and has validated it on welding load/unload and mobile inspection stations. At a 2026 industry event it ran in a NIO welding scenario for three consecutive days — 21.5 hours of cumulative operation — with no errors and no interruptions, by the company's own account. We went deep on this cohort of companies when Unitree went public in Deep Dive — Unitree's 629% debut and the hunt for China's next robot champion; the thesis there was that deployed units, not demo reels, now set valuations. A 21.5-hour run is the unit of evidence that thesis demands.
The case against, which is getting sharper
Two lines of recent work cut directly at the premise. The first is empirical: a August paper introducing WorldEcho finds that today's action-conditioned world models follow expert actions reasonably well but fall apart on off-expert trajectories — they either ignore the commanded action or generate visually invalid rollouts. That is a precise description of the failure you'd expect if a model has learned what futures look like rather than what actions cause. It matters here because the entire point of predicting consequences is to evaluate actions the demonstrator never took. If the world model only works on the distribution it was trained on, the consequence feedback is weaker exactly where it needs to be strongest.
The second is conceptual. A September perspective piece on safety-critical embodied systems argues that high predictive likelihood and visual fidelity do not guarantee a model preserves the evidence needed for safe decisions, and names three structural mismatches: likelihood versus risk, prediction versus intervention, and finite-horizon prediction versus accumulated consequences. Read that against ActEffect and you get a fair critique: ranking three proposals against one real future is an intervention-flavored objective, but it is still trained on demonstrations and still scored by task success, not by whether the robot avoided the outcomes that would have been expensive.
Guangxiang's own researchers concede the ceiling. Benchmarks have to return to real robots and real production lines, they argue, and be judged on task success and the value actually created for customers. That is the right instinct and also a convenient one: the simulation numbers are in, and the real-line numbers are not yet public. The company says Phi-Bot X1 has passed product, functional, and performance validation at formal stations of more than one luxury-brand plant and is entering deployment, with robots expected on the job before the end of the year. Those are the numbers that will settle it.
There is also the land-rush problem. When ten companies ship a world model in four months, some of it is genuine convergence on a necessary component and some of it is a concept being copied because it is the concept of the season. The likely end state is that world models stop being a differentiator and become basic equipment — at which point the differentiation moves back to data, to bodies, and to the unglamorous work of getting a machine to hold a takt time for a year. Related benchmark results are arriving weekly; we noted one in HiDream-O1-Embodied tops RoboColiseum's robustness chart.
What to watch
Three things. First, real-line numbers: whether ActEffect-trained policies hold their LIBERO-PLUS margin on Phi-Bot X1 in a customer's plant, and whether that 21.5-hour run extends to weeks. Second, whether the "train-time-only world model" pattern gets adopted elsewhere — it is a cheap idea to copy and a hard one to credit, so watch whether independent labs reproduce the ablation gaps rather than the headline. Third, whether the off-expert action-following problem gets solved, because until a world model reliably predicts what happens when the robot does something nobody demonstrated, consequence feedback is bounded by the demonstrations it started from.
The deeper bet is worth stating plainly. If robots can learn the physics of consequences instead of the shapes of actions, generalization stops being a data problem and becomes a compression problem — the fiftieth new object is handled with knowledge already extracted from the first ten. That is the promise. A factory floor running unattended for a year is the exam.
If a world model only has to work during training, is it a breakthrough in robot learning — or a clever way to make a benchmark number look like a deployment? Tell us in the comments.
Sources: 智东西 Zhidx — 世界模型不只预测未来，机器人开始争夺"干预世界"的能力 · 量子位 QbitAI — 这个世界模型训练完就"退场"，机器人反而更能干了 · 雷峰网 Leiphone — 光象科技发布物理原生世界模型Phi-WM 1.0 ActEffect · Do Robotic World Models Really Follow Actions? (arXiv 2608.24885) · Rethinking World Models for Safety-Critical Embodied Systems (arXiv 2609.03774) · From World Models to World Action Models: A Concise Tutorial for Robotics (arXiv 2607.00836)
