LingBot-VLA

Robbyant (Ant Group) · 2026-01

VLANew in 2026Not disclosedflow matching

Pretrained on about 20,000 h of real data from 9 bimanual robot types.

Paper

Backbone LLM: Qwen2.5-3B · Qwen2.5-VL-3B config
Vision encoder: Qwen2.5-VL ViT
Action head: MoT flow matching action expert (flow matching)
Chunk: 50 steps
Control rate: Not disclosed
Architecture (original summary): Qwen2.5-VL-3B-Instruct + BAGEL-style Mixture-of-Transformers action expert, flow matching continuous actions, pretraining chunk T=50

Pretraining accelerators: GPU type not disclosed; 261 samples/s in total on 8 GPUs, 8–256 GPU scaling experiments (FSDP, bf16)
Pretraining duration: Not disclosed
Pretraining steps & batch: Not disclosed (scaling experiments local batch 32)
GPU/TPU-hours: Not disclosed
Fine-tuning: GPU and time not disclosed; 130 episodes per GM-100 task, batch 256 for model comparison
Data type: Real-robot teleoperation
Data scale: About 20,000 h, 9 bimanual robot configurations; 39,000 evaluation demonstrations·22,500 trials
Main datasets: In-house bimanual data from 9 embodiments; GM-100
Tasks & evaluation: GM-100 (100 tasks, average over 3 platforms) success rate 17.30% · progress 35.41% (with depth); RoboTwin 2.0 clean 88.56%
Robot: AgiBot G1, AgileX, Galaxea R1Lite·R1Pro, Realman Rs-02, Leju KUAVO 4 Pro, Qinglong, ARX Lift2, Bimanual Franka

Reinforcement learning: Not used · flow matching supervised learning
RL evidence: https://arxiv.org/html/2601.18692v1
Base control: Fixed · Evaluation tasks explicitly stated as tabletop manipulation with chassis and waist fixed (arms, grippers)

Key features: Pretrained on about 20,000 h of real data from 9 bimanual robot types. Released a 3,000→20,000 h data scaling study and the GM-100 benchmark alongside.
Notes: Corrections applied: 261 samples/s is the combined throughput of 8 GPUs (README corrected its 'per GPU' label as an error). arXiv v1 2026-01-26, weights·code 2026-01-27. Apache-2.0.
Verification: Corrections applied