Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training

TL;DR

Xiaomi-Robotics-1 (XR-1) is a VLA foundation model that answers the “what happens if we actually scale a robot policy?” question by pre-training on 100,000 hours of embodiment-free UMI (Universal Manipulation Interface) trajectories spanning 1,700+ scenarios, then post-training on a cross-embodiment mixture including 7,200 hours of in-house real-robot data. The auto-labeling pipeline segments long UMI videos into fixed-length clips and uses a VLM to describe scene-state transitions of grippers and interacting objects, so pre-training supervision is language-conditioned action generation. The blog reports clean pre-training scaling behavior (validation action error drops steadily with data and model size, no saturation), and — crucially — that this pre-training gain transfers to real-robot success in unseen environments after post-training. Downstream: XR-1 reaches 75% average real-robot success at <10 h/task of demos (π0.5 baseline 40%) and 85% at <40 h/task; on four simulation benchmarks it is SOTA, with a +58.3% relative gain over 2nd best on RoboDojo and +23.2% on RoboCasa365.

Key claims

Pre-training on 100,000 hours of embodiment-free UMI trajectories across 1,700+ scenarios (household, commercial, industrial, outdoor) with a VLM-driven auto-labeler that describes gripper + object state transitions per clip, followed by post-training on cross-embodiment robot data + high-quality UMI [§Data, §Method].
Pre-training shows clean scaling behavior: as data and model size grow, validation action error steadily decreases [§Pre-training scaling figure].
Scaling gains from pre-training transfer to post-training: real-robot success rate on unseen environments / unseen object instances rises steadily and predictably as pre-training data and model size grow, with no signs of saturation [§Post-training].
Efficient adaptation to new complex tasks with limited robot data: at an average of <10 h/task, XR-1 hits 75% overall success across phone packing / printer refilling / laundry loading / box packing vs. 40% for the π0.5 baseline at the same budget; increasing to <40 h/task lifts overall success to 85% (vs. π0.5 at 53%) [§Efficient Adaptation table].
State-of-the-art on four simulation benchmarks with the following average success rates and relative gains over 2nd best: RoboCasa 74.5 (+2.6%), RoboCasa365 57.4 (+23.2%), VLABench 59.1 (+11.1%), RoboDojo 13.93 (+58.3%) [§Simulation Benchmarks table].
Two-stage training paradigm: pre-training learns general action-generation representations from UMI; post-training performs embodiment alignment (mapping the general action ability onto real robots via cross-embodiment robot data) and instruction alignment (shifting from “generate actions given a description of state transitions” to “understand a natural-language instruction and execute it directly”) [§Method].

Method

XR-1 follows the LLM-style pre-train + post-train recipe. The pre-training corpus is 100,000 hours of UMI (embodiment-free wrist-camera) trajectories collected across household, commercial, industrial, and outdoor scenes covering 1,700+ scenarios. Manual labeling is infeasible at this scale, so an automatic annotation pipeline splits long videos into fixed-length clips and uses a strong VLM to caption each clip with the state transition it induces on grippers and interacting objects; the model is then trained to generate actions that realize the described state transition, treating language descriptions as the supervision interface.

Post-training runs on a cross-embodiment mixture — in-house real-robot data (7,200+ hours in real homes: sofa-tidying, shoe-cabinet sorting, kitchenware putaway), filtered open-source robot data, and a set of high-quality UMI data with manual temporal segmentation and instruction prompts. Post-training aligns the model along two axes: (a) embodiment alignment maps the pre-trained general action capability onto concrete robot bodies; (b) instruction alignment shifts the interface from state-transition-conditioned generation to instruction-following. Evaluation is deliberately conducted on unseen environments with unseen object instances to test whether pre-training-scale gains survive the post-training stage.

Results

Adaptation efficiency (real robot, average across 4 tasks — Phone Packing, Printer Refilling, Laundry Loading, Box Packing):
- <10 h/task on average: XR-1 = 75%, π0.5 = 40%.
- <40 h/task on average: XR-1 = 85%, π0.5 = 53%. XR-1 hits 100% on Laundry Loading and Box Packing at the higher budget [§Efficient Adaptation table].
Simulation SOTA: XR-1 leads on all four benchmarks: RoboCasa 74.5 (2nd 72.6, +2.6%), RoboCasa365 57.4 (2nd 46.6, +23.2%), VLABench 59.1 (2nd 53.2, +11.1%), RoboDojo 13.93 (2nd 8.80, +58.3%) [§Simulation Benchmarks table].
Scaling curves: pre-training validation action error decreases smoothly with data + model size, and — post-training — real-robot success rate rises smoothly as pre-training data and model size grow, with no observed saturation [§Pre-training scaling figure, §Post-training figure].

Why it’s interesting

XR-1 is the strongest filed data point yet for the scale-of-embodiment-free-data lever in the VLA recipe debate — orthogonal to π*0.6: a VLA That Learns From Experience (RECAP)‘s action-pretraining-then-RL bet and Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models‘s unified-VLM-pointing bet. Where EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data scales 20,854 h of egocentric human video with retargeting, and From Foundation to Application: Improving VLA Models in Practice (LingBot-VLA 2.0)‘s LingBot-VLA 2.0 mixes 10K h of egocentric human + 50K h of teleop robot data at ~60K total, XR-1 scales the UMI wrist-camera substrate to 100K h — a corpus that sits between hand-held human demo and teleop robot data in cost and structure. The predictable pre-train → post-train transfer curve (both axes scale smoothly, no saturation) is the analog of an LLM-style scaling law on the real-robot success axis, going further than EgoScale’s log-linear loss-vs-hours result by showing the gain persists through the embodiment-alignment stage.

The 75% vs 40% headline (vs. π0.5 at <10 h/task) is the sharpest direct comparison to a public baseline: at matched robot-data budget, a stronger pre-trained UMI base ~doubles adaptation success — sharper than Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models‘s clean-teleop lift and more targeted than Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model‘s (same Xiaomi lab family) parallel “world model as data engine” bet, which lifted π0.5 OOD 36.9%→63.2% via synthetic augmentation. Together the two Xiaomi-Robotics releases stake out complementary levers — U0 augments robot data via a generative world model, XR-1 replaces the majority of robot data with embodiment-free UMI + scaling.

For the human-to-robot retargeting cluster: XR-1’s UMI substrate is a different point on the interface-location axis than Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots‘s reduce-to-3-DoF or LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition‘s intent-only recipes — the hand-held UMI gripper is itself the retargeting target, so kinematic + contact match is already baked into the collection tool, and the paper’s contribution is that this shortcut scales cleanly.