Xiaomi-Robotics-0

Xiaomi · 2026-02

VLANew in 20264.7Bflow matching

Learns asynchronous execution with a Λ-shaped attention mask and RoPE position offsets.

Paper

Backbone LLM: Qwen3 · Qwen3-VL-4B config
Vision encoder: Qwen3-VL ViT
Action head: DiT 16-layer flow matching (5 steps) (flow matching)
Chunk: 10 · 4 (SimplerEnv)
Control rate: 30 Hz
Architecture (original summary): Qwen3-VL-4B-Instruct + 16-layer DiT action generator (5 flow matching steps); chunk T=10 (LIBERO·CALVIN)·T=4 (SimplerEnv), 30 Hz timeline

Pretraining accelerators: Not disclosed
Pretraining duration: Not disclosed
Pretraining steps & batch: 40k steps, batch 32,768 (VL:robot sampling 1:6)
GPU/TPU-hours: Not disclosed
Fine-tuning: GPU and time not disclosed; batch 2,048, Lego disassembly 40k · towel folding 80k steps
Data type: Real-robot teleoperation (DROID·MolmoAct·in-house) + web vision-language
Data scale: Robot trajectories about 200M timesteps; VL samples 80M+; in-house task data Lego 338 hours·towel 400 hours
Main datasets: DROID, MolmoAct data, in-house data
Tasks & evaluation: LIBERO 98.7%; SimplerEnv Google Robot 85.5%(VM)·74.7%(VA), WidowX 79.2%; CALVIN ABCD→D 4.75·ABC→D 4.80
Robot: Real-robot bimanual (model not stated in note), simulation Google Robot·WidowX

Reinforcement learning: Not used · Both pretraining and post-training supervised
RL evidence: https://arxiv.org/html/2602.12684
Base control: Fixed · No mention of base (6-DoF bimanual robot)

Key features: Learns asynchronous execution with a Λ-shaped attention mask and RoPE position offsets. Inference latency is 80 ms on RTX 4090.
Notes: arXiv v1 2026-02-13, v2 2026-03-25. Overall scale is in timesteps, so data hours (h) left blank.
Verification: Confirmed