VLA Gallery / All models / New in 2026

Spark-0

Xspark AI·Tsinghua University·MMLab@HKU·HKUST(GZ)·Peking University · 2026-10

VLANew in 2026Not disclosedflow matching

A VTLA model that supervises visual tokens only during training with contact maps from tactile gloves and frozen Depth-Anything-3 features, while a separate tactile pathway conditions only the action expert on per-region force history and 3D taxel states.

Paper

Spark-0: Scaling Spatially Aligned Tactile Pretraining for Robot Dexterous ManipulationarXiv 2610.22342

Overview

Category

New in 2026

Type

VLA

Released

2026-10

Organization

Xspark AI·Tsinghua University·MMLab@HKU·HKUST(GZ)·Peking University

Open weights

Yes

Parameters

Not disclosed (Qwen3-VL-4B + 36-layer width-1024 action DiT; total not stated)

Architecture

Backbone LLM

Qwen3-VL-4B-Instruct

Vision encoder

Qwen3-VL built-in vision (head 384², wrist 256²) + training-only DA3-LARGE geometry teacher (frozen)

Action head

flow matching action expert (Euler 4 steps) → 58-dim bimanual EE pose·hand joints (flow matching)

Chunk

50 steps (RTC replans every 24 steps)

Control rate

Not disclosed

Architecture (original summary)

Qwen3-VL-4B-Instruct VLM (per-layer KV, 64 action queries) → 36-layer width-1024 per-layer flow matching action DiT (even layers VLM cross-attention, odd layers State·Tactile·Action joint self-attention, 24 tactile tokens) → 58-dim (bimanual EE pose 9 + WUJI Hand 2 joints 20) chunk H=50, Euler 4 steps; training-only DA3 geometry·contact map alignment (VLM layer 25)

Training

Pretraining accelerators

96 GPUs (12 nodes × 8, model not stated)

Pretraining duration

Not disclosed

Pretraining steps & batch

400k steps (warmup 20k), global batch 576 (6 per GPU, no accumulation)

GPU/TPU-hours

Not disclosed

Fine-tuning

Per-benchmark downstream fine-tuning 80k steps (warmup 8k), same optimizer·per-GPU batch as pretraining; device count·time not disclosed

Data type

Human video / real-robot teleop / simulation·synthetic

Data scale

Pretraining from 7 sources totaling about 1,944 hours (paper '1.9k h'): HumanTouch 1,000h (100+ tasks·32 operators, 60 Hz)·EgoDex 662h·same-glove robot play 20h·tactile-removed robot play 24h·sim 25h·Fourier ActionNet 153h·AgiBot World 60h; tactile in 44% of samples

Main datasets

HumanTouch; EgoDex; Fourier ActionNet; AgiBot World

Tasks & evaluation

SparkArena-Sim 7 tasks (vision only) average success rate 36.86% (FastWAM 28.29%); SparkArena-Real 5 tasks (with tactile) success rate 52.0%·completion score 62.8 (FTP-1 36.0%·47.8); Ego Humanoid Manipulation Benchmark first on all 8 aggregate metrics (vision only, e.g. Short Seen SR 80.95%)

Robot

Gen 1 Prototype: 2 Marvin M6S CCS-696 arms + 2 WUJI Hand 2 dexterous hands + Moxian tactile gloves (320 taxels per hand), head Orbbec 336·wrist Orbbec 305 cameras

Inference

Inference device & speed

flow matching Euler 4 steps, chunk 50 + training-free RTC (asynchronously requests the next chunk at 26 remaining actions, replans every 24 steps, nominal inference delay 8 steps); policy server is 1 process per GPU (README), GPU model·latency·control Hz not stated

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

1 GPU policy server (model not stated) + websocket robot client

Reinforcement learning & base

Reinforcement learning

Not used · Uses only flow matching imitation learning (action loss) and geometry·contact alignment losses; no mention of RL

Base control

Fixed · Fixed bimanual (Marvin M6S) + dexterous hands, no mobile base

Notes

Key features

A VTLA model that supervises visual tokens only during training with contact maps from tactile gloves and frozen Depth-Anything-3 features, while a separate tactile pathway conditions only the action expert on per-region force history and 3D taxel states. A single checkpoint supports both vision-only and tactile-conditioned deployment, and 1,000 hours of HumanTouch tactile pretraining raises Ego Humanoid vision-only success from 52.5% to 71.7% and SparkArena-Real success from 44.0% to 52.0%.

Notes

Sampling weights HumanTouch 0.35·EgoDex 0.39·same-glove robot 0.09; human data retargeted to WUJI hands via IK; state·action 58-dim (paper) vs 54-dim (README) A/B, RTC replanning interval 24 steps (paper) vs 16 steps (README) A/B; code (GitHub MIT)·pretrained checkpoint pretrain_spark_0 (HF, 400k steps .pt about 10.9GB)·SparkArena released, fine-tuned checkpoints·HumanTouch data release not stated; also integrated into XPolicyLab. Based on: StarVLA (paper [3]) — built on StarVLA·XPolicyLab code (stated in paper)

Verification

Confirmed