Guangxiang Technology and Tsinghua unveil Phi-WM 1.0 ActEffect: a world model that trains robots, then exits the stage

Embodied AI startup Guangxiang Technology, together with Tsinghua University professor Li Shengbo's research group, has released Phi-WM 1.0, a first-generation physics-native world model code-named ActEffect. Its contrarian design: a controlled world model that works only during training, checking what consequences the robot's proposed actions would bring and turning that check into feedback for policy optimization — then exiting the deployment pipeline once training is complete, so the robot no longer unrolls futures or searches candidate actions at execution time.

The company explains that conventional world models act like a mental sandbox a robot carries along: the robot predicts what will happen before deciding how to move, and the more the model thinks, the longer the inference chain grows. ActEffect instead asks the policy to hand over three complete action proposals during training; the controlled world model rolls each one forward and compares the predictions with the real observations in the demonstration data, using a ranking loss to require the fine proposal to be closer to the true future than the coarse one, and the coarse one to beat the original feed-forward proposal.

Two implementation details stand out. The controlled world model ignores task language — it reads only the current visual state and the proposed action, predicting how object positions, poses and scene structure change rather than generating a photorealistic future frame, which is exactly what the team means by physics-native. Gradient truncation on the ranking loss stops bad answers from winning by making themselves worse. When the robot actually starts work, the controlled world model and the future-observation branch are removed together, leaving only the MIP action head to perform coarse proposal and refinement.

On benchmarks, ActEffect reaches 98.8 percent average success on LIBERO, edging DiT4DiT's 98.6 percent; 80.3 percent on LIBERO-PLUS, which injects seven kinds of distribution shift covering camera viewpoint, robot initial state, language phrasing, lighting, background, sensor noise and object layout, versus Fast-WAM's 51.5 percent; and 67.5 percent on RoboCasa-GR1, where a GR-1 humanoid with dual arms, dexterous hands and waist degrees of freedom completes 24 desktop tasks in a 29-dimensional action space, 9.2 points above second-placed ABot-M0. Removing the consequence feedback in ablations drops LIBERO success from 98.8 to 97.0 percent.

Why insist that the world model bow out at all? Guangxiang Technology's answer is an industrial cost ledger: every extra model layer a robot runs in production adds latency, compute and cost, and when a solution is copied across dozens of workstations and hundreds of robots, extra GPUs, power and maintenance are magnified. The real-world logic is simple: keep the extra thinking in training and leave the shorter chain for execution.

Founded in April 2025 and incubated by Tsinghua's School of Vehicle and Mobility and its School of AI, Guangxiang Technology is led by founder and CEO Zhang Tao, a former technical director and head of the spatial-perception engine at Alibaba's AutoNavi, with co-founder Li Shengbo, who has long worked on autonomous driving and embodied intelligence. According to QbitAI's report, the startup has completed real-scenario validation around high-value stations such as welding loading/unloading and mobile inspection, and is in commercial cooperation with several leading domestic and international automakers.

At the 2026 ATC exhibition, the team placed its Phi-Bot X1 robot into a NIO welding loading/unloading scenario, where it ran continuously for three days and logged 21.5 hours with zero errors and zero interruptions. Notably, that record belongs to the whole industrial embodied system — the new model still has to be verified along the same real-world pipeline.

That distinction is the one to watch. The headline scores of Phi-WM 1.0 come from three simulation benchmarks, and factories will not sign off on benchmark numbers alone; acceptance will turn on cycle time, precision, safety and failure rates after delivery. How quickly ActEffect's simulated results transfer to real welding and inspection stations will decide whether the model-first story turns into deployment value.