The most surprising result here is empirical: scaling pre-training to 100k+ hours of real-world manipulation trajectories, combined with automatic natural-language annotation of state transitions, produces a VLA model whose improved pre-training behavior transfers directly into better out-of-the-box real-robot performance in unseen environments. In other words, more real trajectories + language-conditioned action supervision buys practical sim-to-real generalization rather than just higher offline metrics.
Key Findings
- Large-scale real-world pre-training: the paper trains on over 100k hours of manipulation trajectories collected with UMI devices and uses an auto-labeling pipeline to convert clips into language-conditioned action supervision, providing rich behavior-conditioning at scale — this is presented as the main data-side lever. So what: it removes the dependence on handcrafted instruction labels and enables broad, generalizable action generation.
- Two-stage recipe (pre-training → post-training): pre-training imbues broad action capabilities; post-training aligns the model to robot embodiments and imperative human instructions. So what: stronger pre-training models yield better zero-shot and few-shot real-robot performance, reducing downstream fine-tuning needs.
- Empirical scaling and SOTA numbers: Xiaomi-Robotics-1 shows consistent improvements with more data and larger models and establishes new state-of-the-art results on benchmarks — 57.6% success on RoboCasa365 (previous best 46.6%) and an average score of 20.07 on RoboDojo (prior 13.07). So what: concrete, reproducible gains on standard embodied benchmarks indicate practical utility beyond paper claims.
Who it's for and trade-offs
Great fit if you are researching or building real-world robot policies that must follow natural-language instructions and you can access large-scale real trajectory data or want a strong pre-trained policy to fine-tune. Look elsewhere if your constraints are small-data-only or purely simulated datasets (the approach relies on massive real-world trajectories and an auto-labeling pipeline), or if you need guarantees about safety-critical control beyond what imitation-like pre-training provides.
Where it fits
This paper sits at the intersection of embodied AI, multimodal foundation models, and robotics engineering: it emphasizes data scale and language-conditioned action supervision as the primary scaling axis, complementing model-size and post-training alignment. For teams focused on sim-to-real transfer or robot foundation policies, it provides a concrete data+recipe template; for small labs without access to large real trajectory corpora, it signals why shared pre-trained robot policies matter.