The vision language action model drives 20 plus robot bodies from one consumer GPU and triples training data from its January 2026 v1.0 release in six months.
Ant Group's robotics arm Lingbo released LingBot-VLA 2.0 this month as an open-source vision-language-action (VLA) model that drives 20-plus different robot bodies. The release more than triples the team's pre-training corpus, from roughly 20,000 hours in the January 2026 v1.0 release to 60,000 hours now. The pipeline that made that scale jump possible, more than the model name, is the part the rest of physical AI will study.
A VLA maps camera video plus a language instruction to robot motor commands, the same way a language model maps text to text. LingBot-VLA 2.0 trains on 60,000 hours of that input-output pairing, split into 50,000 hours of real-robot trajectory data filtered from a 90,000-hour raw corpus and 10,000 hours of first-person human operation video filtered from 20,000 hours of human footage. The human and robot halves feed the same policy, which is the kind of cross-embodiment bridge that makes a generalist brain plausible. The team positions the 3x scale-up as evidence that robot trajectory data is a renewable input the team can keep compounding, not a one-off labeled asset that has to be hand-curated each release.
The model itself extends VLA control beyond dual-arm manipulation to head, waist, mobile base, and dexterous-hand degrees of freedom, so the same weights coordinate multi-DOF long-horizon tasks rather than only pick-and-place demos. It ships by default with LingBot-Depth, which adds depth heatmaps, object-boundary token masks, and cosine-similarity semantic maps for spatial reasoning. The model also predicts future depth and future semantic features, shifting from reactive frame-by-frame control to anticipatory planning. Release coverage cites sub-130ms inference latency on a single NVIDIA RTX 4090. If that figure survives independent reproduction, v2.0 becomes one of the first open-weights physical-AI policies that runs on commodity hardware instead of a research cluster.
What changed in six months is the data pipeline, not just the architecture. The v1.0 release reached a smaller embodiment set; v2.0 reaches 20-plus, including Leju, Franka, AgileX, ARX Lift2, Galaxea R1Pro, Astribot S1, Unitree G1, Fourier GR-2, and AgiBot A2. The 6B weights live on Hugging Face under the LingBot-VLA 2.0 collection; the code lives on GitHub; and the arXiv technical report documents the data filtering and architecture changes behind the release.
The bet is that robot data follows the trajectory of language-model corpora: each harvest feeds a larger model, which harvests more data, which feeds the next cycle. QbitAI's coverage of the release and the international robotics trade press both lean on the same "universal brain" framing to position v2.0 as a reusable generalist rather than an embodiment-specific specialist.
The open question is whether the scale claim survives contact with independent measurement. The demonstrated tasks in the v2.0 release, including refrigerator loading, cooktop clearing, and condiment organization, are presented as continuous long-horizon video, not as benchmark tables. The arXiv report is where the quantitative numbers live, and per-embodiment fine-tuning recipes have not been published. The sub-130ms RTX 4090 latency figure comes from release coverage and needs community reproduction before it counts as a deployment spec rather than a release-day talking point. Independent reproductions on the 6B weights and the unredacted arXiv benchmark tables are the next test; if they land the way the qualitative demos imply, physical AI starts to look as downloadable as a fine-tuned language model.