VLA Gallery / All models / New in 2026

LingBot-VLA 2.0

Robbyant (Ant Group) · 2026-07

VLANew in 2026About 6.38Bflow matching

Improves on LingBot-VLA, pretraining on 50,000 hours of trajectories from 20 robot configurations plus 10,000 hours of egocentric human video, about 60,000 hours in total, and handles head·waist·mobile base·hand DoF in a 55-dim unified action space.

Paper

From Foundation to Application: Improving VLA Models in PracticearXiv 2607.06403

Overview

Category

New in 2026

Type

VLA

Released

2026-07

Organization

Robbyant (Ant Group)

Open weights

Yes

Parameters

About 6.38B (HF safetensors 6,375,907,511, fp32; checkpoint name 6b; Qwen3-VL-4B backbone + sparse-MoE action expert, active parameters not stated)

Architecture

Backbone LLM

Qwen3-VL-4B-Instruct

Vision encoder

Qwen3-VL built-in vision encoder (not frozen during training)

Action head

sparse-MoE action expert (Qwen2-style, 36 layers, 1 shared + top-4 of 32 routed) flow matching 10 steps (flow matching)

Chunk

50 steps (real-robot deployment example executes 25)

Control rate

Not disclosed (about 130 ms per inference, RTX 4090D)

Architecture (original summary)

Qwen3-VL-4B-Instruct VLM (causal attention) + 2 current·future queries (distilled from LingBot-Depth·DINO-Video) → sparse-MoE action expert (Qwen2-style, MoE in all 36 layers, 1 shared + top-4 of 32 routed) flow matching (10 inference steps) → 55-dim unified state·action (arm·EEF·gripper·hand·waist·head·base) chunk 50

Training

Pretraining accelerators

Not disclosed

Pretraining duration

Not disclosed

Pretraining steps & batch

Not disclosed

GPU/TPU-hours

Not disclosed

Fine-tuning

RoboTwin 2.0 post-training public config: 32 GPUs, batch 1,024, 50k steps, Muon lr 1e-4, L1 flow matching (GPU model·total time not stated); single-node H20 post-training iteration 5.93 s/it (Muon)·4.31 s/it (Distributed Muon); real-robot config batch 256·60k steps; 4×A6000 example needs about 49 GB+ per GPU (README)

Data type

Real-robot teleoperation / human video

Data scale

Pretraining about 60,000 hours: 50,000 hours from 20 robot configurations (filtered from about 90,000 raw hours) + 10,000 hours of egocentric human video (filtered from about 20,000 raw hours); DINO-Video trained on 5M video clips

Main datasets

Self-collected data from 20 robot types; egocentric human video (public and self-collected data, provided by Ant Digital Technologies Phecda Laboratory·Genrobot.ai); GM-100; RoboTwin 2.0

Tasks & evaluation

GM-100 9 tasks generalist (mixed training) progress/success: AgileX Cobot Magic 66.2/34.4 (π0.5 59.1/32.2, LingBot-VLA 1.0 58.2/30.0), Galaxea R1 Pro 34.6/15.6 (π0.5 27.4/8.9); long-horizon mobile manipulation (15 trials): Astribot S1 fridge organizing ID 77.1/60.0·OOD 37.0/13.3, Cobot Magic-ARX X5 stove cleaning ID 84.3/66.7·OOD 67.5/40.0 (all above π0.5); RoboTwin 2.0 clean 93.52%·randomized 92.80%

Robot

Evaluation: AgileX Cobot Magic, Galaxea R1 Pro, Astribot S1, Cobot Magic-ARX X5; 20 pretraining robots: Franka, Flexiv Rizon 4, AgileX, ARX Lift2, UR7e, AgiBot G1, Galbot G1, Moz1, Realman Rs-02, Galaxea R1Pro·R1Lite, Astribot S1, Zerith H1, Leju KUAVO 4 Pro, Tienkung, QingLong, MagicBot Gen1, Unitree G1, Fourier GR-2, AgiBot A2

Inference

Inference device & speed

About 130 ms per inference on RTX 4090D (10 denoising steps, torch.compile; real-robot deployment executes 25 steps of the chunk); public checkpoint validation uses FP32 inference, BF16 uses less memory but success rate may differ (README); depth·video teachers not used at inference

Inference memory

Weights released in fp32 (about 6.38B); validation uses FP32 inference, BF16 saves memory (figure not stated)

Weights bf16 (calc.)

12.8 GB

Where it runs

RTX 4090D inference server (README example)

Reinforcement learning & base

Reinforcement learning

Not used · flow matching supervised learning (pretraining·post-training); depth·video feature distillation is an auxiliary loss

RL evidence

https://arxiv.org/html/2607.06403

Base control

Wheeled (Direct base velocity/pose output) · Directly outputs the 3-dim locomotion signal of the unified action vector (labeled base.velocity in the LeRobot conversion) chunked together with arm·waist·head; long-horizon tasks evaluated on Astribot S1·Cobot Magic-ARX X5 mobile bases (wheeled per robot specs; the paper only says 'mobile base')

Notes

Key features

Improves on LingBot-VLA, pretraining on 50,000 hours of trajectories from 20 robot configurations plus 10,000 hours of egocentric human video, about 60,000 hours in total, and handles head·waist·mobile base·hand DoF in a 55-dim unified action space. Adds sparse MoE to the action expert and learns future prediction as an auxiliary task by distilling current·future queries from LingBot-Depth·DINO-Video.

Notes

Subtask annotations auto-generated with Qwen3.6-27B (vocabulary of 18 atomic actions); human video MANO hand trajectories recovered via SLAM·hand pose estimation; ablations report relative joint actions over absolute joints, MeanStd normalization, and L2 (MSE) flow matching work better; pretraining uses loss-free MoE load balancing (DeepSeek-V3-style bias), post-training config uses sequence auxiliary loss·z-loss; code·weights (robbyant/lingbot-vla-v2-6b, RoboTwin post-trained lingbot-vla-v2-6b-robotwin) Apache-2.0, LeRobot conversion (lerobot/lingbot_vla_v2_base) available; arXiv v1 2026-07-07. Based on: A Pragmatic VLA Foundation Model (arXiv 2601.18692) — improves on the previous LingBot-VLA (stated in paper)

Verification

Confirmed