Source-linked AI summary
Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories
Xiaomi Robotics Team, Jun Guo, Piaopiao Jin, Jason Li, Peiyan Li, Yingyan Li, Futeng Liu, Wanli Peng, Optimus Qin, Yifei Su, Nan Sun, Qiao Sun, Runze Suo, Heyun Wang, Yunhong Wang, Rujie Wu, Caoyu Xia, Lina Zhang, Jack Zhao, Guoliang Chen, Wenlong Chen, Xinze He, Bin Li, Qing Li, Zhuorong Li, Heng Qu, Wenxuan Song, Diyun Xiang, Yifan Xie, Peiran Xu, Hangjun Ye, Wen Ye, Han Zhao, Quanyun Zhou
TL;DR
Robot learning is constrained by limited, redundant, and narrow trajectory data. Xiaomi-Robotics-1 addresses this with large-scale pre-training and post-training, achieving state-of-the-art results across four simulation benchmarks, including 57.4% on RoboCasa365.
Problem
Robot learning lacks scalable, diverse trajectory data because teleoperated datasets are difficult to scale, redundant, and concentrated on limited tasks and environments.
Method
Xiaomi-Robotics-1 uses two-stage pre-training and post-training on over 100,000 hours of trajectories with scalable language auto-labeling of scene-state transitions.
Results
Xiaomi-Robotics-1 achieves state-of-the-art results across four simulation benchmarks, including 57.4% success on RoboCasa365, 10.8 percentage points above previous best methods.
Takeaways & Limitations
Increasing pre-training data improves out-of-the-box performance in unseen environments, supporting further scaling as a promising direction for robot learning.
Takeaways & Limitations
RoboDojo evaluation omits history observations, yielding lower memory performance than a model that explicitly models memory.
Abstract
from arXiv · showhide
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. Project page: https://robotics.xiaomi.com/xiaomi-robotics-1.html
1 Introduction
Xiaomi-Robotics-1 is a scalable VLA foundation model trained with a two-stage pre-training/post-training recipe on over 100k hours of real-world trajectories. It improves with data and model scale, transfers those gains to unseen-environment robot performance, and achieves strong simulation and data-efficient adaptation results.
- Motivation: Xiaomi-Robotics-1 addresses limited teleoperated-data diversity by combining heterogeneous real-world, cross-embodiment, and vision-language data at scale.The approach is motivated by teleoperated trajectories being redundant and concentrated in narrow task and environment distributions.
- Method: The model uses pre-training on over 100k hours of UMI manipulation trajectories followed by post-training that aligns general action representations with robot embodiments and imperative instructions.The recipe uses scalable, diverse data for pre-training and high-quality cross-embodiment data for post-training; trajectory clips receive state-transition language prompts.
- Scaling: Xiaomi-Robotics-1 scales effectively: larger data and models reduce pre-training validation action errors, and stronger pre-training models yield higher unseen-environment real-robot success after post-training.This transfer demonstrates that pre-training scaling behavior carries directly into post-training performance.
- Results: 57.6% on RoboCasa365 and 20.07 on RoboDojo establish new state-of-the-art results over 46.6% and 13.07, respectively, while Xiaomi-Robotics-1 leads all four evaluated simulation benchmarks.It also achieves a 75% average success rate across four dexterous tasks using fewer than 10 hours of data per task, versus 40% for π0.5 [5].
- Capabilities: The model performs diverse tasks in unseen environments out-of-the-box and adapts efficiently to novel downstream tasks with minimal data.Post-training enables language-instruction prompting across a wide range of tasks, while the model autonomously completes a room-level suitcase-packing task lasting over 10 minutes.
3 Experiments
Experiments show that Xiaomi-Robotics-1 scales with pre-training data and model size, and that these gains transfer to out-of-the-box performance in unseen environments. It also achieves state-of-the-art benchmark results and efficient adaptation to challenging novel tasks with limited data.
- Pre-Training Scaling: Validation action error decreases as data scale increases, while model scaling consistently improves action-prediction precision but produces smaller gains than data scaling.The model-scaling result suggests that billion-parameter capacity may already be sufficient for the current dataset distribution.
- Post-Training Scaling: 75% overall success with full pre-training data rises monotonically from 26% without action pre-training, with the final data-scale increase adding 6 percentage points in unseen environments.Post-training evaluations use unseen environments and object instances, showing that pre-training data scaling transfers to out-of-the-box real-robot performance.
- Post-Training Scaling: 79% overall success for the 10B model exceeds 75% for 5B and 61% for 2B, with the strongest gains on shoe tidying.These variants use the same 20k hours of UMI pre-training data, demonstrating that model scaling also improves out-of-the-box real-robot performance after post-training.
- Data-Efficient Adaptation: 75% average success and 90% average progress with less than 10 hours per task outperform π0.5’s 40% success and 66% progress across four held-out novel tasks.The tasks cover bimanual coordination, long-horizon mobile manipulation, deformable-paper handling, and multi-object language grounding; the advantage is attributed to large-scale pre-training and post-training alignment.
- Benchmark Comparisons: 57.4% average success on RoboCasa365 surpasses previous methods by 10.8 percentage points, while Xiaomi-Robotics-1 achieves state-of-the-art results across all four challenging benchmarks.It also reaches 74.5% average success in RoboCasa and performs strongly on the most challenging Composite-Unseen split.
4 Related Work
Related work frames robot learning around scaling laws and robot foundation models, particularly world-action and vision-language-action paradigms. Existing VLA research advances reasoning, action representations, and heterogeneous pre-training, while this work focuses on scaling behavior in robot foundation models.
- Scaling for Robot Learning: Scaling robot learning follows evidence from language and multimodal foundation models that increasing data, compute, and model capacity improves performance predictably [22] [27].Recent advances in large language models [72] and multimodal foundation models [1] [2] [3] [63] motivate applying scaling to robot learning.
- Robot Foundation Models: Robot foundation models use large-scale datasets to generalize across diverse environments and adapt efficiently to novel downstream tasks.
- Robot Foundation Models: Recent VLA research improves robot learning through intermediate reasoning representations [13] [19] [34] [64] [84] [89] [93], learned action tokenizers and flow matching [4] [48] [58] [82], and heterogeneous multi-embodiment pre-training [8] [30] [54] [69] [82].
- Robot Foundation Models: This work follows the VLA paradigm while addressing the complementary question of how robot foundation models scale.
5 Conclusions
Xiaomi-Robotics-1 is a foundational VLA model that performs diverse mobile manipulation tasks in unseen environments without task-specific adaptation and efficiently adapts to challenging novel tasks with minimal data. Its broad manipulation capabilities are learned from over 100,000 hours of real-world trajectories during pre-training.
- Xiaomi-Robotics-1 follows instructions to perform a wide range of mobile manipulation tasks out-of-the-box in unseen environments.
- The model efficiently adapts to novel challenging tasks using a minimal amount of data.
- Pre-training on over 100,000 hours of real-world manipulation trajectories gives Xiaomi-Robotics-1 broad and generalizable manipulation capabilities.
Contributions
The authors are listed in alphabetical order.
- The paper’s authors are listed alphabetically.