Robbyant Open-Sources LingBot-VLA 2.0 Across 20 Robot Morphologies
Robbyant, the embodied-AI unit inside Ant Group, has open-sourced LingBot-VLA 2.0, a vision-language-action model built to run on many robot bodies instead of one fixed arm stack. The company says the model was pre-trained on about 60,000 hours of physical data and covers 20 robot configurations from 17 manufacturers.
What shipped
According to Robbyant’s GitHub release and secondary coverage from IT Brief Asia, the pre-training mix is roughly 50,000 hours of cleaned robot trajectories plus 10,000 hours of egocentric human manipulation video. Embodiments span single-arm, dual-arm, half-humanoid, humanoid, and mobile bases — including hardware from Leju, AgiBot, Unitree, AgileX, Galaxea, Galbot, RealMan, Franka, ARX, Fourier, Flexiv, and others listed in the release materials.
The action interface is a 55-dimensional canonical vector that can carry arm joints, end-effector pose, gripper, hand joints, waist, head, and mobility signals rather than dual-arm-only control. Architecture notes on the repo include sparse MoE layers in the action expert and dual-query distillation from LingBot-Depth and DINO-Video teachers for geometry and future-scene cues.
Benchmarks and pilots
On Robbyant’s reported numbers for the GM-100 bimanual benchmark, LingBot-VLA 2.0 posted higher progress / success than π0.5 and GR00T N1.7 on AgileX Cobot Magic (66.2 / 34.4 vs lower baselines) and competitive results on Galaxea R1 Pro. Long-horizon mobile tasks on Astribot S1 and Cobot Magic-ARX X5 also beat π0.5 in the company’s tables for both in-domain and out-of-domain settings. On RoboTwin 2.0 simulation, the release lists 93.52% (clean) and 92.80% (randomized) success averages.
Deployment notes claim about 130 ms inference on an RTX 4090-class card with 10 denoising steps after post-training. IT Brief Asia reports commercial pilots in retail sorting, logistics, and industrial automation with partners including Leju, Ti5 Robot, GuoDa Drugstore, and Longsheng Technology.
Weights and code are public under Apache-2.0 on GitHub and Hugging Face (lingbot-vla-v2-6b), with a RoboTwin post-training checkpoint dated 2026-07-25. Technical report: arXiv:2607.06403.
A Human’s Take
I’m here for models that treat “which robot am I on?” as a first-class problem instead of a footnote. Cross-morphology VLA work is how smaller teams stop rebuilding brains every time they change a chassis. The pilot list is still thin on independent third-party metrics — when warehouses publish shift-level numbers, that’s when this graduates from open-source moment to labor tool.