VLA Gallery / All models / Mobile manipulation

Xiaomi-Robotics-1

Xiaomi (Xiaomi Robotics Team) · 2026-07

VLAMobile manipulation5.1Bflow matching

A two-stage VLA pretrained on 100K+ hours of UMI trajectories with VLM auto-labels (scene state-transition descriptions), then aligned to robots and instructions with about 10K hours of cross-embodiment data.

Paper

Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World TrajectoriesarXiv 2607.15330

Overview

Category

Mobile manipulation

Type

VLA

Released

2026-07

Organization

Xiaomi (Xiaomi Robotics Team)

Open weights

Yes

Parameters

5.1B (5B public release: VLM 4.4B + DiT 604M); scaling variants 2.6B (2.1B+470M)·10.5B (8.8B+1.5B)

Architecture

Backbone LLM

Qwen3-VL-4B-Instruct (5B public release; 2B·10B variants use VLM 2.1B·8.8B)

Vision encoder

Qwen3-VL built-in vision encoder

Action head

DiT flow matching (Euler 5 steps, adaLN time conditioning) + VLM Choice Policies auxiliary regression (flow matching)

Chunk

30 steps × 60-dim (public code default; H not stated in paper)

Control rate

Not disclosed

Architecture (original summary)

Qwen3-VL (Qwen3-VL-4B-Instruct for the 5B public release) + small-hidden DiT with the same layer count, combined as Mixture-of-Transformers → VLM KV cache·proprioception-conditioned flow matching (Euler 5 steps) → relative delta EE pose·gripper·waist·base velocity chunk (public code default 30 steps × 60-dim); VLM trained with Choice Policies auxiliary regression (K candidates·scores)

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed (auto-labeling alone about 2 weeks)

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Post-training (about 10K hours, VL:open-source:UMI instruction:in-house = 0.5:0.5:0.5:8.5) accelerator·time not disclosed; downstream 4 tasks 144 hours (low-data 36 hours) with the Xiaomi-Robotics-0 asynchronous training recipe, GPU·steps not disclosed (public demo config batch 48)

Data type

UMI handheld demos / real-robot teleop / open-source robot data / web vision-language

Data scale

Pretraining on 100K+ hours of UMI trajectories (scaling experiments use about a 20K-hour subset); post-training about 10K hours (7,200+ hours in-house robot, 1,000+ hours instruction-labeled UMI, Bridge V2·RT-1·DROID); downstream 144 hours

Main datasets

In-house UMI data (private); in-house mobile·bimanual robot data; Bridge V2; RT-1; DROID

Tasks & evaluation

RoboCasa 74.5%; RoboCasa365 57.4% (57.6% in the introduction, A/B; prior best ABot-M0.6 46.6%); VLABench average SR 59.1·PS 70.3·IS 69.9; RoboDojo score 20.07·success rate 13.93%; real-robot post-training zero-shot evaluation, 4 tasks (unseen environments) 2B 61%·5B 75%·10B 79%; downstream 4 tasks low-data 75% (π0.5 40%)·high-data 85% (π0.5 53%)

Robot

In-house wheeled bimanual mobile manipulator (waist lift) and fixed bimanual robot (model name not stated); 3 cameras ego·left/right wrist; sim RoboCasa·RoboCasa365·VLABench·RoboDojo

Inference

Inference device & speed

flow matching 5-step Euler, supports asynchronous execution (action prefix); latency·Hz·device not stated (README 'optimized for consumer GPUs')

Inference memory

Not disclosed

Weights bf16 (calc.)

10.2 GB

Where it runs

Inference server (GPU model not stated) – client architecture, asynchronous execution

Reinforcement learning & base

Reinforcement learning

Not used · Pretraining·post-training·fine-tuning all imitation learning (flow matching + Choice Policies regression + NTP); no mention of RL

RL evidence

https://arxiv.org/abs/2607.15330

Base control

Wheeled (Direct base velocity/pose output) · The VLA directly outputs base velocity (3-dim) and waist position delta in the unified action vector, chunked together with arm EE actions

Notes

Key features

A two-stage VLA pretrained on 100K+ hours of UMI trajectories with VLM auto-labels (scene state-transition descriptions), then aligned to robots and instructions with about 10K hours of cross-embodiment data. Zero-shot performance in unseen environments rises with pretraining data and model scale, reaching RoboCasa365 57.4% and a RoboDojo score of 20.07, above the prior best (46.6%·13.07).

Notes

Pretraining objective is flow matching + Choice Policies regression + 0.1×NTP (VL:UMI = 1:9); the DiT is blocked from seeing the VLM's action-token KV to cut the copying shortcut; state-transition captions auto-generated with Qwen3.5-27B; 10+ minute room-scale suitcase-packing demo; code·checkpoints released 2026-08-03 (Apache-2.0, HF Xiaomi-Robotics-1-5B model_states.pt 10.2 GB, per-benchmark RoboCasa·RoboCasa365·VLABench checkpoints), 2B·10B not released, pretraining UMI data not released; arXiv v2 2026-07-22. Predecessor: Xiaomi-Robotics-0 (arXiv 2602.12684) — inherits the asynchronous training recipe and VL data (stated in paper)

Verification

Confirmed