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

1 Introduction

• Pre-training validation action errors fall as data and model scale increase, and stronger pre-training models yield higher out-of-the-box success in unseen-environment real-robot evaluations.The reported scaling behavior transfers from pre-training to post-training.Source: §1 · p. 2
• Xiaomi-Robotics-1 averaged 75% success across four challenging tasks with less than 10 hours of data per task on average, versus π0.5 [5] at 40%.It is reported to complete a long-horizon, room-level suitcase-packing task lasting over 10 minutes.Source: §1 · p. 3
• 57.6% success on RoboCasa365 (previous best: 46.6%) and 20.07 average score on RoboDojo (prior state-of-the-art: 13.07) set new state-of-the-art results across four simulation benchmarks.The model also autonomously completed a room-level suitcase-packing task lasting over 10 minutes.Source: §1 · p. 3

2.1 Model

• The Mixture-of-Transformers combines a pre-trained VLM and DiT for flow-matched action chunks; VLM Choice Policies add candidate-action supervision, while DiT excludes action-related tokens from attention.The paper says excluding those tokens mitigates a shortcut in which the DiT could copy VLM-generated actions instead of grounding its own generation in visual and textual context.Source: §2.1 · pp. 3–4

2.2.1 Pre-training

• Pre-training uses over 100,000 hours of UMI trajectories from diverse environments; Qwen3.5-27B captions fixed-length segments’ gripper and object state transitions, and a producer–consumer pipeline labels the corpus in roughly two weeks.The model learns to generate actions that move the scene from its observed state to the language-described target state.Source: §2.2.1 · pp. 4–5

2.2.2 Post-training

• Post-training aligns UMI action generation to robot embodiments and state-transition captions to imperative prompts, using about 10,000 hours including over 7,200 robot hours, over 1,000 instruction-labeled UMI hours, and open-source data.The robot data includes mobile manipulators and dual-arm robots, while the instruction-labeled UMI data uses human-style prompts.Source: §2.2.2 · p. 6
• Post-training unifies end-effector frame orientations across robot and UMI data so similar motions yield consistent action values across hardware, masks missing action dimensions, and enables out-of-the-box unseen-environment instruction following and efficient adaptation.The action representation also uses base velocity and relative waist-position changes for mobile robots.Source: §2.2.2 · pp. 6–7

3.1 Pre-training: Data and Model Scaling

• On held-out validation data, action error decreases as pre-training data increases; 50% and 100% data yield monotonic loss declines, while 12.5% and 25% settings show overfitting.Larger models also improve action-prediction precision, but their gains are less pronounced than those across data scales, suggesting data volume is the primary bottleneck at billion-parameter scale.Source: §3.1 · pp. 7–8

3.2.1 Effectiveness of Scaling Pre-training Data

• Across four tasks in unseen environments and with unseen objects, out-of-the-box success increases with pre-training data, from 26% without action pre-training to 75% with all 20k hours.Using 12.5% of the data more than doubles baseline success (26% versus 53%); doubling data from 50% to 100% adds 6 percentage points, with no sign of saturation.Source: §3.2 · p. 8, §3.2.1 · pp. 8–9, Figure 7 · p. 9

3.2 Post-training: Out-of-the-Box Evaluation in Novel Environments

• Post-training evaluation tests four tasks seen in training but uses unseen environments and object instances, measuring performance out-of-the-box without per-task or per-environment fine-tuning.The tasks are shoe storage, bag packing, table organization, and sofa tidying.Source: §3.2 · p. 8, Figure 7 · p. 9

3.2.2 Effectiveness of Scaling Model Size

• Overall out-of-the-box success rises from 61% for the 2B model to 75% for 5B and 79% for 10B, while the shoe-tidying rate reaches 92% for 10B.Across three of four tasks, performance improves with model size; combined with data scaling results, the authors identify data scale and model size as complementary axes.Source: §3.2.2 · pp. 9–10, Figure 8 · p. 10

3.3 Downstream Fine-tuning: Efficient Adaptation to New Tasks

• With less than 10 hours per task, Xiaomi-Robotics-1 reaches 75% average success and 90% progress, versus π0.5 [5] at 40% and 66%, and raises printer-refilling success from 20% to 70%.The four held-out tasks span bimanual coordination, long-horizon mobile manipulation, deformable paper handling, and multi-object language grounding; the authors attribute data efficiency to large-scale pre-training and post-training alignment.Source: §3.3 · pp. 10–11, Figure 10 · p. 12

3.4 Simulation Benchmarks

• Xiaomi-Robotics-1 achieves average success rates of 74.5% on RoboCasa and 57.4% on RoboCasa365, outperforming previous best methods by 10.8 percentage points, with its largest improvement on Composite-Unseen.Source: §3.4 · p. 13
• On VLABench, Xiaomi-Robotics-1 leads in average success and progress; on RoboDojo, it beats the second-best method by 7% in average score and 5.13% in average success rate.It ranks first on four of five RoboDojo dimensions, but scores below Hy-Embodied-0.5-VLA [85] on memory because evaluation omits history observations.Source: §3.4 · pp. 13–14, Table 4 · p. 14

Scaling for Robot Learning

• Robot-learning scaling is constrained by costly teleoperation and narrow data diversity; UMI devices enable diverse in-the-wild collection, and this work uses over 100k hours to study pre-training and post-training scaling.Prior work also converts egocentric human manipulation trajectories into robot data through representation alignment or motion retargeting [15] [40] [50].Source: §4 · p. 14

Robot Foundation Models

• World-action models predict observations or dynamics to guide action, whereas vision-language-action models use pre-trained vision-language backbones for semantic action prediction.WAM research includes visual planning, joint future prediction and action generation, geometric structure, and heterogeneous video-action training; VLA research builds on rich semantic knowledge [4] [5] [8] [9] [10] [24] [25] [30] [47] [54].Source: §4 · pp. 14–15
• Recent VLA advances target reasoning, action representation, and large-scale heterogeneous pre-training; this work instead investigates scaling through massive trajectory datasets with precise language annotations.The approaches described include embodied reasoning tokens and visual chains of thought, learned tokenizers and flow matching, and training across robot embodiments to improve transfer and downstream performance.Source: §4 · p. 15

• Xiaomi-Robotics-1 achieves strong state-of-the-art results across four challenging simulation benchmarks.The conclusion does not provide benchmark-specific scores.Source: §5 · p. 16
• The model follows instructions for diverse mobile manipulation tasks in unseen environments and adapts to novel challenging tasks with minimal data.Source: §5 · p. 16
• Pre-training uses over 100,000 hours of real-world manipulation trajectories, with an auto-labeling pipeline generating language prompts that describe scene state transitions.Post-training aligns pretrained capabilities with robot embodiments and imperative prompts using a cross-embodiment dataset.Source: §5 · p. 16
• Performance improves with increasing pre-training data scale and model size, and this scaling transfers to out-of-the-box performance in unseen environments after post-training.Source: §5 · p. 16