A Beijing embodied-AI startup has open-sourced a robot world model that runs on RGB alone, and it is topping benchmark tables built by researchers who assumed depth was mandatory.
Wujie Dongli published MWA-Phoenix, a fully pretrained 7B parameter latent world model designed for on-device deployment. The company reports leading results across six evaluations covering environment perturbation, task generalisation, dexterous manipulation and real-machine execution.
The benchmark detail matters more than the model size. MWA-Phoenix takes RGB input only and uses no depth sensor. It ranks first on LIBERO-Plus with a 94.63 per cent overall success rate, first on LIBERO-Pro at 66.48 per cent, and reaches 82.67 per cent average task success on the RoboCasa GR1 tabletop benchmark, a 7.50 percentage point gain over its predecessor.
What the six benchmarks say
Cross-embodiment work is where most world models quietly fail, so the numbers there are the interesting ones. On Bench2Dex the model averages 45.6 per cent success. On RoboTwin 2.0, which tests two-arm coordination, it reaches 72.42 per cent average success.
Wujie Dongli also placed it on the RoboDojo real-machine leaderboard across three robot bodies and 18 tasks, which is the closest thing in the report to evidence of physical execution rather than simulated policy quality.
Teaching contact through touch
Fine manipulation is where vision-only systems break, so the team shipped a second module. Phoenix-ILA encodes synchronous motion and tactile sequences together, letting contact participate directly in action learning rather than being inferred from pixels.
The gains on real tasks are concrete. Cucumber peeling rose from 20 per cent to 40 per cent success, and plate wiping from 75 per cent to 90 per cent once the module was added.
A third release, Phoenix-PhysCanvas, is an interactive simulation world model built to supply reinforcement learning with an environment it can actually push against. After post-training inside that framework, physical success rates on a test-tube task rose from 58.33 to 75 per cent, and on a block-stacking task from 80 to 86.67 per cent.
The architecture and the data moat
Under the hood, MWA-Phoenix uses DINOv3 as its visual backbone, fuses geometric priors with temporal information through dual-teacher distillation, and folds environment dynamics prediction and action generation into a single joint learning framework.
The company says it holds more than 200,000 hours of multimodal data and is training larger models on it, aimed at dexterous tasks in commercial spaces and industrial lines.
That data reserve is the part Western robotics teams should read twice. Open weights lower the cost of entering the model layer. They do not transfer the operational data that keeps the model improving after release, and the Chinese firms publishing weights fastest are also the ones collecting that data at scale.
Open source is not generosity in this category. It is a distribution strategy wrapped around a data advantage.
Editor’s note: This is an adapted translation of the original Gasgoo report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.gasgoo.com/apps/e060bad9eca91db7f01403fc7120041f/robot/news/70475549-%E6%97%A0%E7%95%8C%E5%8A%A8%E5%8A%9B%E5%8F%91%E5%B8%83%E5%B9%B6%E5%BC%80%E6%BA%90mwa/.