VLA Gallery / All models / New in 2026

τ0-VLA (tau-0)

Shanghai Innovation Institute·Agibot Finch·CUHK · 2026-08

VLANew in 2026About 3.0Bflow matching

A test-time computation hierarchical VLA whose high-level policy, with execution memory, uses a world model to predict outcome videos of candidate subtasks and a value model to score them for beam search.

Paper

τ0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time ComputationarXiv 2608.16885

Overview

Category

New in 2026

Type

VLA

Released

2026-08

Organization

Shanghai Innovation Institute·Agibot Finch·CUHK

Open weights

Yes

Parameters

About 3.0B (low-level policy, HF safetensors BF16 3,015,404,136; backbone Qwen3.5-2B); high-level policy is 3 Qwen3.5-9B-based models + Step1X-Edit world model, system total not stated

Architecture

Backbone LLM

Qwen3.5-2B (low-level policy VLM; high-level policy is Qwen3.5-9B)

Vision encoder

Qwen3.5-2B built-in vision (multi-view RGB)

Action head

MoT action expert conditional flow matching (masked, Euler 10 steps) → 40-dim chunk (flow matching)

Chunk

30 steps × 40-dim (LIBERO checkpoint 10, 8 executed)

Control rate

About 30 Hz control loop (high-level subtask updated about every 1 s)

Architecture (original summary)

Qwen3.5-2B VLM + Mixture-of-Transformers action expert (per-layer joint attention) → conditional flow matching (masked, Euler 10 steps) → 40-dim unified state·action (EEF Rot6D·gripper·waist·planar base velocity·arm joints) chunk H=30, about 30 Hz control; high-level policy (Qwen3.5-9B proposal·value·reflection models + Step1X-Edit world model) generates subtask language commands

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

Stage-3 per-task fine-tuning (small per-task demos) accelerator·time not disclosed; LIBERO checkpoint 60,000-step post-training (GPU not stated); public example config bf16·DeepSpeed ZeRO-1·lr 2e-5·1,000 steps

Data type

Real-robot teleop / autonomous policy rollouts / UMI data / multimodal VL co-training

Data scale

40,115 hours (internal about 23.4K hours: AGIBOT G1 21.9K·G2 585·ARX AC One 578·Franka 347; public data about 16.7K hours, of which open-source UMI 9.25K) + multimodal VL data (scale not stated)

Main datasets

Internal AGIBOT G1/G2·ARX·Franka demos; open-source UMI data; other public robot datasets (names not stated)

Tasks & evaluation

Real-world 4 long-horizon tasks (10 trials per task, AGIBOT G1): average SR 27.5%·progress 80.1%, hierarchical (Plan Once) 45.0%·87.85%; baselines π0.5 22.5%, GR00T N1.7 2.5%, LingBot-VLA 0.0%; with TTC Make Milk Tea 5→7/10, Book Organization 6→9/10, Clean Room 5→7/10; Collect Laundry 10/10 (π0.5 9/10); LIBERO average 97.35% (README)

Robot

AGIBOT G1 (wheeled humanoid, omnidirectional 4-wheel-steering base, 7-DoF dual arms·parallel grippers); ARX AC One (6-DoF X5 dual arms); bimanual Franka Research 3 (two 7-DoF arms); training includes AGIBOT G2·UMI data

Inference

Inference device & speed

Low-level policy executes chunks in an about 30 Hz control loop; high-level policy updates the subtask about every 1 s via an asynchronous worker; inference GPU·latency not stated (LIBERO example server environment RTX 4090, 8 of 10 predicted actions executed)

Inference memory

Not disclosed

Weights bf16 (calc.)

6 GB

Where it runs

Inference GPU not stated (LIBERO example server RTX 4090)

Reinforcement learning & base

Reinforcement learning

Not used · No mention of RL in the paper·model card; low-level policy uses flow matching imitation learning, high-level models use supervised fine-tuning, training data includes autonomous policy rollouts

Base control

Wheeled (Direct base velocity/pose output) · The 40-dim unified action includes 2-dim planar base velocity (23–24) → the policy directly drives the AGIBOT G1 omnidirectional wheeled base (Clean Room·Collect Laundry, etc.); ARX·Franka tasks have no base

Notes

Key features

A test-time computation hierarchical VLA whose high-level policy, with execution memory, uses a world model to predict outcome videos of candidate subtasks and a value model to score them for beam search. The low-level policy trained on 40,115 hours of real data reaches an average SR of 27.5% on 4 long-horizon tasks (π0.5 22.5%), and the hierarchical system (Plan Once) raises it to 45.0%.

Notes

Low-level policy 3-stage training (KI co-training → end-to-end co-training → per-task adaptation); Gemma4-31B-it used for high-level data labeling; public HF weights are the low-level pretrained policy (not a zero-shot controller, for post-training), high-level Proposal·World Model released separately (2026-09); HF release 2026-07-27, arXiv v1 2026-08-17; code·weights Apache-2.0, example data CC BY-NC-SA 4.0

Verification

Confirmed