Wujie Dongli Open-Sources World Model MWA-Phoenix, Claims Top Spot on Six Benchmarks

Chinese company Wujie Dongli (无界动力) open-sourced MWA-Phoenix on October 22. It is a foundation world model that predicts environment dynamics in latent space rather than in pixels. According to the developers, the model took first place on six benchmarks: LIBERO-Plus, LIBERO-Pro, RoboCasa GR1, Bench2Dex, RoboTwin 2.0, and the real-robot RoboDojo leaderboard. The six platforms cover simulation, physical robots, and several types of tasks.

Six leaderboards that test different things

Few teams claim leadership across such a broad set of benchmarks. Usually a team optimizes a model for one or two tests and leaves the rest aside. Here, the choice reads as a bid for generality: each platform tests a different layer of the problem.

• LIBERO-Plus and LIBERO-Pro: robustness to environmental perturbations and transfer to reworded tasks and changed task conditions.
• RoboCasa GR1: household scenarios in kitchen scenes with a humanoid platform.
• Bench2Dex: complex object manipulation, where grasping dexterity matters.
• RoboTwin 2.0: bimanual tasks with domain randomization.
• RoboDojo: a testbed with real robots, where simulator tricks don't help.

In the release description, Wujie Dongli highlights four axes: reliability under disturbances, generalization after a task change, handling of complex objects, and transfer between different robot bodies. These are the places where manipulation policies break most often. A robot trained with one camera and one lighting setup is often thrown off by a shift in the image. A model learned on one arm cannot control another.

Why latent space suits robots

World models for robotics fall into two camps. One camp generates future frames, which is expensive and noisy: most pixels are irrelevant to deciding where to move a manipulator. The other predicts how a compressed representation of the scene evolves and works with that. MWA-Phoenix belongs to the second camp. The latent representation discards background, textures, and glare, keeping geometry, object positions, and the relations between them. Such a scene depends less on changes in lighting or camera, and the action policy receives a compact input. In theory, this yields both faster inference and a more direct path to transfer between robots, because general knowledge of physics is separated from specific kinematics.

This is exactly the argument the company puts in the shop window. If a foundation model really does separate scene understanding from the robot's body, then adapting to a new platform comes down to light fine-tuning rather than collecting a dataset from scratch.

Open code as a way to check the claims

Leaderboards tend to flatter the picture: configurations are tuned, run protocols differ, and public rankings are sometimes held for weeks. An open release matters because the claim can be rechecked. Any lab can run the model on its own testbed and see whether the result holds outside the home benchmarks.

Wujie Dongli is also entering a market where foundation models for robots are becoming infrastructure. Physical AI has already grown into a separate field, and Chinese teams are betting on openness to build a community and data around their models faster.

For those building robots or researching the field, the checklist is practical. First, reproduce the results on LIBERO-Plus and RoboTwin 2.0 on your own infrastructure: it is a cheap test of whether the numbers are honest. Then run the model on your own platform with minimal fine-tuning and verify the claimed transfer between bodies. Finally, read the license and the contents of the open artifacts: weights, training code, and data do not always ship together. If transfer between robots holds up, teams without a large fleet of machines will get a ready-made foundation instead of months of collecting demonstrations.