Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories

Abstract

We present Xiaomi-Robotics-1, a foundational vision-language-action (VLA) model capable of (1) following diverse language instructions to perform a wide range of mobile manipulation tasks in unseen environments out-of-the-box, and (2) efficiently adapting to novel downstream tasks with minimal fine-tuning data. We propose a two-stage training recipe consisting of pre-training and post-training. During pre-training, we imbue the model with broad and generalizable action-generation capabilities by training on over 100k hours of real-world manipulation trajectories collected via UMI devices. Crucially, we develop a scalable auto-labeling pipeline that annotates trajectory clips with natural languages describing scene state transitions, providing rich and precise conditioning for action learning. During post-training, we aim to align these capabilities with robot embodiments and imperative instructions that humans naturally use to prompt robots. Extensive experiments demonstrate strong scaling behavior. Xiaomi-Robotics-1 consistently improves with increased data scales and model sizes during pre-training. This scaling behavior directly transfers to post-training, where a stronger pre-training model yields better out-of-the-box real-robot performance in unseen environments. Furthermore, Xiaomi-Robotics-1 serves as a strong robot foundation policy that can be efficiently fine-tuned on complex, dexterous tasks with high data efficiency. Across multiple simulation benchmarks, Xiaomi-Robotics-1 outperforms state-of-the-art methods. Notably, it establishes a new state-of-the-art with a 57.4% success rate on RoboCasa365, surpassing the previous best of 46.6%. Furthermore, it achieves an average score of 20.07 on RoboDojo, significantly outperforming the prior state-of-the-art (13.07). Code and model checkpoints will be released.

Key Contributions

• A scalable auto-labeling pipeline that annotates raw UMI-collected trajectory clips with natural-language descriptions of scene state transitions — turning cheap teleoperation-style capture into richly-conditioned training data without manual annotation.
• A two-stage pre-training → post-training recipe: pre-training builds broad, generalizable action-generation capability; post-training aligns the model to specific robot embodiments and the imperative phrasing humans actually use when commanding robots.
• Empirical demonstration that VLA scaling laws hold and have not saturated at 100k+ hours — data and model-size increases both keep improving downstream performance, and pre-training gains transfer cleanly into post-training / fine-tuning.
• Evidence that a strong pre-trained foundation policy can be fine-tuned on complex, dexterous downstream tasks with high data efficiency (i.e., the foundation model does the heavy lifting; task-specific data requirements shrink).

Methodology

Training data was collected via UMI (Universal Manipulation Interface) devices — a portable, robot-free demonstration-capture method — at a scale of 100k+ hours of real-world manipulation trajectories. The auto-labeling pipeline segments trajectories into clips and generates natural-language descriptions of state transitions, which condition the action-generation model. Pre-training optimizes for broad action-generation capability across this corpus; post-training aligns the model with specific robot embodiments and natural imperative instruction phrasing. The model was evaluated across multiple simulation benchmarks (RoboCasa365, RoboDojo) and on out-of-the-box real-robot deployment in unseen environments.

Results

• RoboCasa365: 57.4% success rate, vs. prior best of 46.6%.
• RoboDojo: average score of 20.07, vs. prior state-of-the-art of 13.07.
• Scaling behavior is consistent and unsaturated across both data volume and model size during pre-training, and this scaling advantage transfers into post-training / real-robot performance.
• Fine-tuning on complex, dexterous downstream tasks shows high data efficiency, attributed to the strength of the pre-trained foundation policy.

Limitations

Not explicitly enumerated in the abstract; the paper notes code and model checkpoints "will be released" (not yet public at time of ingest), so independent verification of the reported numbers is not yet possible from public artifacts alone.

Full Content

This is a Xiaomi corporate-research release (33-author team) rather than a university lab paper, positioning Xiaomi as a serious entrant in the VLA foundation-model race alongside NVIDIA (GR00T), Physical Intelligence (π-series), and NVIDIA/Google DeepMind. The headline claim — that scaling to 100k+ hours of real-world data shows no saturation — is the most consequential data point for MenFem's embodied-AI-as-compute-demand thesis: it argues the ceiling on useful training data for physical-world foundation models is still far off, meaning data-collection infrastructure (UMI-style capture devices, auto-labeling pipelines) becomes as strategically important as the compute used to train on it. The auto-labeling pipeline is itself notable: it removes the classic bottleneck (manual language annotation of demonstration data) that has throttled VLA dataset growth relative to LLM text-corpus growth.

Source: Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories by Jun Guo et al., Xiaomi Robotics Team