VLA Gallery / All models / New in 2026

Qwen-VLA

Qwen Team · 2026-05

VLANew in 2026Not disclosedflow matching

Attaches a DiT flow-matching decoder to a Qwen VLM, unifying manipulation·navigation·trajectory prediction in one action·trajectory space, with per-robot prompt conditioning so one set of weights handles multiple embodiments.

Paper

Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot EmbodimentsarXiv 2605.30280

Overview

Category

New in 2026

Type

VLA

Released

2026-05

Organization

Qwen Team

Open weights

No

Parameters

Not disclosed (Qwen3.5-4B backbone + action expert about 1.15B; no total stated)

Architecture

Backbone LLM

Qwen3.5-4B

Vision encoder

Qwen3.5-4B built-in ViT (spatial merging)

Action head

DiT flow-matching decoder (about 1.15B, 16 blocks) (flow matching)

Chunk

16 steps (8 navigation waypoints)

Control rate

Per dataset (specified by prompt)

Architecture (original summary)

Qwen3.5-4B (native multimodal, ViT + spatial merging) + single-stream DiT flow-matching action decoder about 1.15B (16 blocks×70.8M, AdaLN time conditioning) → manipulation action chunk H=16, 8 navigation waypoints (Δx,Δy,Δθ; sliding-window execution in VLN-CE), few Euler steps; robot tag·arm configuration·control frequency·chunk size conditioned via text prompt

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

T2A 2,000 steps (final pipeline default); later-stage steps·batch not disclosed (in T2A ablation, peaks at 2,000 steps·plateaus up to 10,000 steps)

GPU/TPU-hours

Not disclosed

Fine-tuning

SFT·ALOHA fine-tuning accelerator·time not disclosed; RL (PPO+GAE, RLinf, SimplerEnv): per iteration 8 rollout epochs × 128 env steps (8,192 chunks), actor lr 5e-6·value lr 1e-4

Data type

Real-robot teleoperation / human video / simulation·synthetic / web vision-language

Data scale

10,000+ hours of public manipulation data + 1,000+ hours of in-house real-robot trajectories; 8M+ synthetic sim trajectories; T2A language-action synthesis about 7.2M trajectories·14,000+ hours; egocentric human video (EgoDex 829 hours, EgoVerse 1,300+ hours, etc.); mixing ratio manipulation 74.2%·human 6.0%·navigation 7.5% (instruction 4.3·object search 2.3·tracking 1.0)

Main datasets

RoboMIND; DROID; BridgeData V2; RT-1; BC-Z; AgiBot World; Ego4D·EPIC-KITCHENS(VITRA); EgoDex; EgoVerse; navigation data (source not stated)

Tasks & evaluation

LIBERO 97.9%; RoboCasa-GR1 56.7%; Simpler-WidowX 73.7%; RoboTwin 2.0 Easy 86.1%·Hard 87.2%; R2R-CE val-unseen OS 69.0·SR 57.5·SPL 51.2; RxR-CE SR 59.6·SPL 47.8; SimplerEnv-OOD 32.0%; DOMINO zero-shot SR 26.6%; real ALOHA average 83.6% (OOD 76.9%)

Robot

Real robot ALOHA bimanual (fixed); pretraining WidowX·Google Robot·Franka·ARX5·Fourier GR-1·Mobile ALOHA·AgiBot A2-D·Galaxea R1·AIRBOT MMK2·TienKung (arm actions only); navigation is a VLN-CE sim 3-DoF planar agent

Inference

Inference device & speed

Not disclosed (only states it targets low-latency real-time control with few Euler steps)

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Used (VLA RL post-training) · PPO+GAE (RLinf) from multi-task SFT checkpoint, SimplerEnv sparse binary success reward

RL evidence

https://arxiv.org/html/2605.30280v2

Base control

Sim only (Waypoints → tracking controller) · No real mobile base control: manipulation uses fixed arms (real ALOHA etc.); navigation executes 8 relative waypoints in a sliding window on a VLN-CE sim 3-DoF planar agent

Notes

Key features

Attaches a DiT flow-matching decoder to a Qwen VLM, unifying manipulation·navigation·trajectory prediction in one action·trajectory space, with per-robot prompt conditioning so one set of weights handles multiple embodiments. A single generalist model scores LIBERO 97.9%·R2R-CE SR 57.5%·RxR-CE SR 59.6% and averages 83.6% on real ALOHA (π0.5 71.6%).

Notes

No real mobile base control experiments: navigation is evaluated only in VLN-CE R2R/RxR sim and the real robot is a fixed ALOHA, so as instructed it is listed as 'New in 2026'·base 'Fixed'. Paper byline 'Owen Team' (README: 'Qwen Team'). GitHub serves as an info·issue channel with no checkpoint links (as of 2026-10). Total parameters·training compute·inference latency not disclosed. Data hours exist only per source, so the total cell is left empty.

Verification

Corrections applied