Training-Only World Model Improves Robot Success Rates Across Benchmarks

The method tries to answer a long-standing question in embodied AI: whether a robot that has learned actions from demonstrations can also learn from the consequences of those actions to improve its policy. ActEffect aims to solve this without burdening real-time deployment. In a production environment, every extra inference step of a world model creates additional latency, computing and costs; keeping the model in training saves that overhead while still forcing the policy to absorb outcome-based experience.

To generate outcome feedback, the policy is first asked to produce three complete versions of an action. The first comes from a feed-forward branch, which acts like the robot's first reaction. The second is a coarse proposal from an MIP action head. The third is a refined version of that coarse proposal and becomes the final executed action. The three proposals are complete in the sense that they can be directly put into the world model, unlike the noisy intermediate samples of diffusion or flow-matching policies. Then a controlled world model receives the current visual state and each proposal, and predicts how the scene will evolve. The language instruction is not fed to the world model; it remains with the VLA policy, because the physical outcome of pushing an object depends on the object's position, the environment and the applied action, not on the wording of the task. ActEffect also encodes future states in the frozen DINOv3 visual feature space instead of generating a photorealistic future image, capturing changes in object positions, poses and scene structure while preserving the physical-native focus.

The three predictions are compared with the real future observations contained in the training data. The model requires the refined proposal to be closer to the true future than the coarse proposal, and the coarse proposal to be closer than the feed-forward first response. A gradient-truncation mechanism on the ranking loss prevents poor answers from improving their ranking simply by becoming worse. The feedback signal is written into the policy parameters, so once training is completed, the world model and future-state branch are removed, and only the MIP action head remains to handle coarse proposal and refinement during execution.

On the LIBERO benchmark, ActEffect achieved an average success rate of 98.8 percent, slightly higher than DiT4DiT's 98.6 percent, while retaining the robot's basic operation ability. On LIBERO-PLUS, which adds camera-view, initial-state, language, lighting, background, sensor-noise and object-layout shifts, it recorded 80.3 percent, compared to 51.5 percent for Fast-WAM. On RoboCasa-GR1, which requires a GR-1 humanoid robot to perform 24 tabletop manipulations in a 29-dimensional action space, the average success rate was 67.5 percent, 9.2 percentage points higher than ABot-M0 and 10.8 points higher than Fast-WAM. The larger action space and more complex body structure made the consequence feedback more effective. Ablation studies also confirmed the role of each component: removing the consequence feedback dropped LIBERO accuracy from 98.8 percent to 97.0 percent, replacing DINOv3 features with VLM representations resulted in 97.3 percent, and removing the ranking loss brought the score down to 98.1 percent.

The training-only usage is closely connected to industrial deployment. When a robot operates in an auto plant, repeated future unfolding would be multiplied across dozens of workstations and hundreds of devices, adding real GPU, power and maintenance costs. ActEffect extracts an extra layer of supervision from existing demonstration data and leaves a shorter inference chain for execution.

According to the report, Guangxiang Technology has carried out real-site verification around welding loading/unloading, mobile inspection and other high-value automotive workstations, and has commercial cooperation with several leading automakers at home and abroad. At the 2026 ATC exhibition, its robot Phi-Bot X1 ran for three days in a NIO vehicle welding loading/unloading scenario, accumulating 21.5 hours with zero failures and zero interruptions. The report notes that this 21.5-hour record came from the complete industrial embodied system, and the new ActEffect model still has to be verified along the real-robot chain.

The company was founded in April 2025 and jointly incubated by Tsinghua University's School of Vehicle and Mobility and School of Artificial Intelligence. CEO Zhang Tao, a Tsinghua Ph.D. and a Politecnico di Milano postdoc, previously served as technology director of Alibaba's AutoNavi and led its spatial-perception engine. Co-founder Li Shengbo, a Tsinghua professor, has long worked on autonomous driving and embodied-intelligence research with industrial applications. The company builds robot hardware, data and algorithm systems alongside the model, which it sees as necessary for turning benchmark success into production-line reliability.