Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model

TL;DR

Xiaomi-Robotics-U0 is a 38B decoder-only autoregressive multimodal model that treats embodied generation as a continuation of foundation image/video pretraining, jointly optimizing text-to-image, image editing, multi-view embodied scene generation, controllable embodied transfer, and embodied video rollout in a single next-token loss. Built on Qwen3-32B + EMU3.5 with IBQ as the shared image tokenizer, it is continually trained in two stages so that the pre-trained visual prior is preserved rather than overwritten by robot data. It reports state-of-the-art results across single-step and sequential embodied benchmarks, beats GPT-Image-2.0 in human evals of embodied scene generation / transfer, ranks first on WorldArena for embodied video generation, and — critically — lifts the OOD success rate of a π₀.₅ policy from 36.9% to 63.2% on real-world manipulation when its outputs are used as a data engine. A companion FlashAR+ inference scheme delivers up to 82.9× speedup and drops 1024² single-sample latency from 450.77s to 5.44s.

Key claims

• 38B autoregressive multimodal model with IBQ as a shared tokenizer, jointly optimizing T2I, image editing, embodied scene generation, embodied transfer, and embodied video generation in a single next-token loss [Abstract].
• A two-stage continual-training strategy keeps the pre-trained foundation image-generation capability intact while adapting to robot-centric reasoning and interaction modeling [Abstract, §Method].
• Beats GPT-Image-2.0 in most human-eval comparisons on Capability I (multi-view embodied scene generation), with clear multi-view-consistency wins at comparable instruction-following [§Capability I].
• Disentangles embodied scenes into five independently controllable dimensions, enabling exponentially many robot-scene variants from compact descriptions while preserving robot pose and manipulation state [§Capability II].
• Ranks 1st among 100+ models on WorldArena for embodied video generation with EWMScore_P 73.64, ahead of SisyphusWorld (73.06), BWM-Fast (72.71), and SACWM (72.67) [WorldArena table].
• Improves the out-of-distribution success rate of a π₀.₅ policy from 36.9% → 63.2% on challenging real-world manipulation when U0-generated style-transfer data is added to real demonstrations for post-training [Abstract].
• FlashAR+ inference (diagonal-parallel decoding + vLLM paged KV-cache batch scheduling + acceleration adaptation) yields up to 82.9× generation speedup and reduces 1024×1024 single-sample latency from 450.77s to 5.44s [§FlashAR+].

Method

Xiaomi-Robotics-U0 is a 38B decoder-only multimodal autoregressive model built on Qwen3-32B and continually trained from EMU3.5, using IBQ as a discrete image tokenizer so text, single-view images, multi-view images, and video frames all share one token space and one next-token objective. Continual training runs in two complementary stages: the first preserves foundation image-generation capability by mixing T2I / image-editing data with embodied data, the second adapts the model to robot-centric objectives (multi-view scene generation, structured embodied transfer, and video rollout from a single observation). Embodied scenes are explicitly disentangled into five independently controllable dimensions so that transfer edits (e.g. lighting, tablecloth, background) recombine compositionally without disturbing robot pose or interaction state.

At inference, FlashAR+ extends the FlashAR autoregressive-decoding acceleration scheme to X2I (anything-to-image) tasks: a lightweight post-training + acceleration adaptation is combined with vLLM’s paged KV cache and batch scheduling, plus diagonal-parallel decoding over the image-token grid. The result is native support for multi-reference image editing under the same acceleration path used for scene generation and video rollout.

Results

On single-step embodied generation, U0 outperforms GPT-Image-2.0 in most head-to-head human evaluations for both Capability I (multi-view embodied scene generation, dominant wins on multi-view consistency) and Capability II (controllable embodied transfer, stronger transfer robustness while matching instruction-following) [§Capability I, §Capability II]. On sequential embodied generation, U0 ranks first on WorldArena among 100+ models with EWMScore_P 73.64, leading on visual quality (61.78), 3D accuracy (92.05), controllability (91.60), interaction quality (87.30), and instruction following (93.86) [WorldArena table]. As a data engine for downstream policies, adding U0-generated style-transfer data to real demonstrations for π₀.₅ post-training moves OOD manipulation success from 36.9% to 63.2% under held-out tablecloths and low/colored/patterned lighting [Abstract, §Policy Post-Training]. FlashAR+ delivers up to 82.9× end-to-end inference speedup and reduces 1024×1024 single-sample latency from 450.77s to 5.44s while preserving generation quality [§FlashAR+].

Why it’s interesting

This is the strongest concrete argument yet for the “foundation image/video model as the base of the world model” bet: rather than train a world model from robot trajectories with limited data (as Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation does for Qwen-RobotWorld, or HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds does for 3D assets), U0 explicitly continues pretraining from EMU3.5 (Emu3.5: Native Multimodal Models are World Learners) and demonstrates that the pretraining prior is what enables both multi-view consistency and the 26-point OOD policy lift. That last number is the load-bearing evidence: it turns the “world models as scalable data engines” thesis from an aspirational claim into a measured downstream metric on π₀.₅ (see π0.7: A Steerable Robotic Foundation Model with Emergent Compositional Generalization for the base policy). FlashAR+‘s 82.9× speedup also puts autoregressive image/video generation back on the serving Pareto frontier against diffusion — a live tension for Diffusion serving optimization and worth watching alongside TileRT: Tile-Based Runtime for Ultra-Low-Latency LLM Inference.

See also

• World Foundation Models — U0 is the first filed WFM built by continual training on a foundation UMM base rather than from-scratch on robot data.
• Unified Multimodal Models — extends the UMM paradigm (T2I + editing + generation) with embodied scene/video generation as first-class tasks under one next-token loss.
• VLA Models — evaluated as a data engine for a VLA (π₀.₅), not as a VLA itself; 36.9% → 63.2% OOD lift is the headline transfer number.
• Synthetic Training Data — style-transfer augmentation over held-out backgrounds/lighting is the concrete mechanism for the OOD policy lift.
• Emu3.5: Native Multimodal Models are World Learners — the direct EMU3.5 ancestor U0 continues training from.
• Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation — parallel language-conditioned WFM, contrasts with U0’s multi-view + transfer emphasis.
• Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models — Qwen-RobotManip also argues alignment/pretraining prior must precede robot scale.
• π0.7: A Steerable Robotic Foundation Model with Emergent Compositional Generalization — π₀.₆/π₀.₅ family, whose OOD success rate is the downstream metric moved by U0-generated data.
• Autoregressive Video Generation — FlashAR+ inference for AR image/video generation revives the serving-throughput argument.