XiaomiRobotics/Xiaomi-Robotics-1

TL;DR: Xiaomi-Robotics-1 (XR-1) is an open-weights Vision-Language-Action model pre-trained on 100K+ hours of embodiment-free UMI (handheld gripper) trajectories and post-trained on 10K+ hours of cross-embodiment robot data — probably the largest open robot-policy scaling effort to date. The interesting claim is not the benchmark table but the recipe: they sidestep the robot-data bottleneck by auto-labeling human-collected UMI clips with a VLM and training on "make the scene go from state A to described state B," then aligning to real robots and natural-language instructions afterward — an explicit LLM-style pretrain/post-train split, with claimed no-saturation scaling.

What it is and why it matters

The central bet in robot foundation models is whether scaling data buys generality the way it did for language. XR-1 is a strong datapoint for "yes, if you decouple data collection from robots." UMI devices (handheld grippers with wrist cameras) let humans collect manipulation trajectories at high throughput with no robot in the loop; Xiaomi collected 100K+ hours across 1,700+ scenarios (household, commercial, industrial, outdoor). For comparison, most open VLA efforts (OXE-scale) sit in the low thousands of hours of heterogeneous robot data. The team claims pre-training scaling behavior transfers through post-training to real-world performance "with no signs of saturation" — the kind of claim worth verifying in the tech report (arXiv:2607.15330).

How it works

Architecture: a pre-trained Qwen3-VL backbone coupled to a Diffusion Transformer action head via a Mixture-of-Transformers — the DiT mirrors the VLM's layer count but with a smaller hidden size, so action denoising is cheap relative to the VLM forward pass. Asynchronous execution hides inference latency for real-time control. This is broadly the π0/GR00T family design; the architecture is not the novelty.

The novelty is the data pipeline and training split:

Stage 1 — embodiment-free pre-training. A VLM auto-labeling pipeline segments long UMI trajectories into fixed-length clips and generates descriptions of each clip's state transition. The model learns to generate actions that take the current scene to the described target state. This makes annotation nearly free at 100K-hour scale — no human labels, no robots.

Stage 2 — post-training alignment, along two axes: embodiment alignment (mapping general action generation onto actual robot kinematics via in-house robot data, filtered open-source robot data, and manually annotated UMI data) and instruction alignment (shifting from scene-transition descriptions to executing natural-language commands directly).

Evidence

SOTA claimed on four simulation benchmarks, with the biggest gains where generality matters (RoboCasa365's 365-task spread, RoboDojo):

Note the shape: RoboCasa is a wash (+1.9 points — within noise for sim evals), while RoboCasa365 (+10.8 pts) and RoboDojo (+58% relative, but 13.9% absolute — everyone fails this benchmark) show the pretraining paying off on breadth. The README also claims XR-1 substantially beats π0.5 on few-hours-of-demos adaptation to new real-world tasks, but the table is truncated in the README I was given — no numbers to report, check the tech report.

What's actually released, and caveats

Released: post-training/inference/eval code (Apache-2.0), checkpoints on Hugging Face and ModelScope including per-benchmark fine-tuned variants, and eval guides for RoboCasa/RoboCasa365/VLABench. Built on HF transformers pinned to 4.57.1, claimed to run on consumer GPUs.

Not released, as far as the README states: the 100K-hour UMI dataset, the auto-labeling pipeline, and pre-training code. So you can fine-tune and deploy their model, but not reproduce the interesting part.

Things to scrutinize: (1) "unseen environments" in real-world evals — with 1,700+ pretraining scenarios, what counts as unseen deserves inspection; (2) how much the transfer depends on the target robot using a UMI-style parallel gripper — cross-embodiment alignment on 10K hours may not stretch to dexterous hands; (3) sim benchmark checkpoints are per-benchmark fine-tunes, not one generalist checkpoint; (4) "no signs of saturation" is a strong claim resting on the paper's scaling curves, not anything in the repo.

Trying it

No quickstart snippet in the README. Shortest path:

git clone https://github.com/XiaomiRobotics/Xiaomi-Robotics-1
pip install transformers==4.57.1
# checkpoints: https://huggingface.co/collections/XiaomiRobotics/xiaomi-robotics-1
# sim eval, e.g.: follow eval_robocasa365/README.md with Xiaomi-Robotics-1-RoboCasa365

Project page with demos: https://robotics.xiaomi.com/xiaomi-robotics-1.html · Paper: arXiv:2607.15330.