Xiaomi Robotics
4 min
Recent foundation image and video generation models offer strong generalization and controllability, but their direct application to embodied scenarios is limited by requirements for multi-view consistency, geometric coherence, and robot embodiment constraints. Existing methods typically adapt foundation models with limited robot data, often sacrificing visual knowledge acquired during large-scale pre-training. We present Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive model for unified embodied synthesis. It treats embodied generation as an extension of foundation image and video generation and jointly optimizes text-to-image generation, image editing, embodied scene generation, embodied transfer, and embodied video generation. This unified framework preserves the generalization of the pre-trained world foundation model while adapting it to embodied settings. Xiaomi-Robotics-U0 is the first model to support high-quality multi-view scene generation across multiple robot embodiments and to introduce structured, controllable embodied transfer for fine-grained editing while preserving multi-view consistency and interaction dynamics. It achieves state-of-the-art results on single-step and sequential generation tasks, outperforming GPT-Image-2.0 in human evaluations of embodied scene generation and transfer, ranking first on World Arena for embodied video generation, and improving the out-of-distribution success rate of pi_0.5 from 36.9% to 63.2% on challenging real-world manipulation tasks. These results show that foundation world models can serve both as embodied world models and scalable data engines for embodied intelligence. Code and checkpoints are available at https://robotics.xiaomi.com/xiaomi-robotics-u0.html.
How can foundation models, typically trained on massive internet-scale visual data, be adapted for embodied intelligence without sacrificing their general-purpose visual reasoning and generation capabilities? The authors seek to bridge the gap between general image synthesis and the strict requirements of robotics, such as multi-view consistency, geometric coherence, and temporal interaction dynamics.
The authors introduce Xiaomi-Robotics-U0, a 38-billion-parameter model that treats embodied generation as a natural extension of general-domain image and video synthesis. By utilizing a unified autoregressive objective, the model is trained on a mixture of general-domain data (text-to-image, image editing) and embodied data (multi-view scene generation, manipulation videos). The framework employs a structured control mechanism that disentangles scene elements—such as workspace, objects, and lighting—allowing for fine-grained, controllable generation. To handle sequential tasks, the model learns interleaved subtask-subgoal sequences, effectively transforming it into a scalable data engine that can generate diverse, coherent interaction trajectories for downstream policy learning.
Xiaomi-Robotics-U0 achieves state-of-the-art performance across both single-step and sequential generation tasks. It outperforms GPT-Image-2.0 in human evaluations for embodied scene generation and transfer and ranks first on the World Arena benchmark for embodied video generation. Crucially, the model demonstrates practical utility by improving the out-of-distribution success rate of robot manipulation policies (pi_0.5) from 36.9% to 63.2%. These results suggest that unified foundation world models can serve as both effective simulators and scalable data generators for robotics.
This work provides a path toward "agentic" data generation, where robots can autonomously synthesize diverse, physically plausible experiences to train themselves. By unifying general visual knowledge with robot-specific constraints, the model overcomes the limitations of smaller, specialized embodied datasets, offering a scalable solution for training robust embodied intelligence.
Alex: So the model isn't just generating pretty pictures. It's generating synthetic experience.
Sam: That's a fair way to put it. And the results reflect that. In testing, the success rate for tasks in unfamiliar, previously unseen environments improved from roughly 37% to over 63% when this synthetic data was used. That's a meaningful gain, and it came without the robot ever having physically visited those spaces.
Alex: The paper also mentions something called "FlashAR+." What does that do?
Sam: Speed is the issue it solves. Normally, generating an image pixel by pixel, one step at a time, is slow—too slow for a robot that needs to react in real time. FlashAR+ changes the approach by generating groups of pixels simultaneously, working along a diagonal pattern across the image rather than going one by one. Think of it like the difference between one artist painting a canvas alone versus a whole team each working on a different section at the same time. The result is much faster, which matters enormously when the robot needs to respond to something happening right now.
Alex: That makes sense. Real-time decisions can't wait for a slow image pipeline to catch up. Are there any trade-offs the authors flag?
Sam: They're candid about the costs. Balancing three very different types of data—text, images, and physical robot actions—in a single training process requires significant computing resources. It's a complex, ongoing engineering challenge, and the authors don't present it as fully solved.
Alex: So where does that leave us? This feels like a meaningful step, but not the end of the road.
Sam: That's the right read. The study suggests we're getting better at giving robots a richer, more reliable understanding of physical space—without having to manually collect enormous amounts of real-world data to do it. The approach is promising, but the authors are measured about it. There's still real work to be done before this scales cleanly to the full messiness of everyday environments.
Alex: It's a useful reminder that progress in robotics tends to be incremental rather than sudden. Thanks for walking us through it, Sam, and thanks to everyone listening to ResearchPod.