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

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

Yingyan Li
2026.07.26
·Arxiv·by JunhaRyu
#Foundation Model#Robotics#Scaling#Trajectory#VLA

Key Points

  • 1Xiaomi-Robotics-1 is a foundational vision-language-action model trained on over 100k hours of real-world trajectories using a novel two-stage recipe that combines broad pre-training with cross-embodiment post-training.
  • 2The model demonstrates strong scaling properties, where increased pre-training data and model size directly lead to superior performance in unseen environments and enhanced zero-shot task execution.
  • 3Experimental results confirm state-of-the-art performance across multiple simulation benchmarks and demonstrate the model's ability to achieve high success rates on complex, dexterous tasks with minimal fine-tuning.

Xiaomi-Robotics-1 is a foundational vision-language-action (VLA) model designed to improve robot generalization through large-scale, two-stage training. It combines a pre-trained vision-language model (VLM) with a diffusion transformer (DiT) to perform diverse mobile manipulation tasks and adapt efficiently to novel downstream scenarios.

Methodology and Architecture

The model adopts a Mixture-of-Transformers (MoT) architecture. A pre-trained Qwen3-VL encodes visual observations and language instructions, while a parallel DiT generates action chunks via flow matching. To accelerate convergence, the VLM uses a "Choice Policy" mechanism, where multi-layer perceptron (MLP) heads predict candidate action chunks and associated scores. To prevent the DiT from merely copying these VLM-generated actions, the architecture explicitly excludes action-related tokens from the DiT’s attention computation.

The flow-matching training objective for the DiT is defined as:
LFlow(θ)=∣∣vθ(ot,l,st,a~t:t+Hτ,τ)−u(a~t:t+Hτ,at:t+H,τ)∣∣22L_{Flow}(\theta) = ||v_\theta(o_t, l, s_t, \tilde{a}^\tau_{t:t+H}, \tau) - u(\tilde{a}^\tau_{t:t+H}, a_{t:t+H}, \tau)||^2_2
where a~t:t+Hτ=τat:t+H+(1−τ)ϵ\tilde{a}^\tau_{t:t+H} = \tau a_{t:t+H} + (1-\tau)\epsilon represents the noisy action. The VLM supervision adds a regression loss:
LRegression(θ)=∣∣a^t:t+H∗−at:t+H∣∣1+∑k∣∣s^k−sk∣∣22L_{Regression}(\theta) = ||\hat{a}^*_{t:t+H} - a_{t:t+H}||_1 + \sum_k ||\hat{s}_k - s_k||^2_2
where s^k\hat{s}_k is the predicted score for the kk-th candidate action chunk.

Two-Stage Training Recipe

  1. Pre-training: The model is trained on over 100,000 hours of real-world UMI (Universal Manipulation Interface) trajectories. Because manual annotation is infeasible at this scale, the authors implemented a scalable auto-labeling pipeline using a VLM to generate language descriptions of scene state transitions. This imbues the model with broad, generalizable action-generation capabilities.
  2. Post-training: The model is aligned to specific robot embodiments and imperative human instructions using 10,000 hours of cross-embodiment data (including static arms and mobile manipulators). This phase transitions the model from state-transition logic to task-oriented control.

Scaling and Performance

The paper demonstrates consistent scaling behavior: pre-training validation action error (MSE) decreases as both data volume and model parameters increase. This scaling property transfers to post-training, where models pre-trained on larger datasets exhibit significantly better out-of-the-box success rates in unseen environments.

In comparative benchmarks, Xiaomi-Robotics-1 establishes new state-of-the-art results:

  • RoboCasa365: 57.4% success rate (vs. 46.6% prior SOTA).
  • RoboDojo: Average score of 20.07 (vs. 13.07 prior SOTA).
  • Fine-tuning efficiency: The model achieves 75% success across four complex dexterous tasks using less than 10 hours of data per task, significantly outperforming existing foundation policies like π0.5\pi0.5.