Ant Group's Lingbo open sourced LingBot VLA 2.0, a vision language action model, across 17 manufacturers' hardware with shared weights; vendor benchmarks put it ahead of Physical Intelligence's π0.5 and NVIDIA's GR00T N1.7.
Ant Group's Lingbo team has open-sourced a robot-control model that runs across hardware from 17 manufacturers with one set of weights. It is the latest public test of the embodied AI field's bet that robot AI should stop being built one-brain-per-body.
The release, LingBot-VLA 2.0, is a vision-language-action model, a class of AI that ingests camera input and language instructions and outputs physical motion commands for a robot arm, base, or humanoid. Ant Lingbo describes it as a full upgrade of LingBot-VLA 1.0, first open-sourced in January, with pretraining now drawn from roughly 60,000 hours of curated real-robot trajectories (up from about 50,000 in v1.0), spliced with first-person human-operation data, the kind of body-control signal a purely teleoperated corpus does not contain.
The release lands on the central fault line in embodied AI. The field has built around per-body policies because joints, sensors, and dynamics differ enough that a model trained on one body barely transfers. Lingbo's bet is the version of that argument the LLM era already won: train one generalist base, post-train it cheaply onto each new hardware, and pay the per-body cost only at the end.
VLA 2.0 ships with weights claimed to drive 20 configurations spanning single-arm, dual-arm, wheeled, and bipedal platforms. For the first time in the Lingbo line, the model is reported to output commands for more than the end-effector: head, waist, and mobile-base joints, the joints that let a humanoid or a mobile manipulator do more than wave a gripper. The OEM list is long and includes publicly shipping Chinese robots from Unitree (宇树), AgiBot/Zhiyuan (智元), Galbot (银河通用), Astribot (星尘), AgileX (松灵), Realman (睿尔曼), Fourier (傅立叶), and Franka, alongside embodied-AI startups such as Leju, RoboSpace, Ark, X-Humanoid, MagicLab, Qianxun, Zero-form, Flexiv, and Qinglong.
'Supported' here is a data-compatibility claim, not a coalition of endorsers. A robot body is on the list because Lingbo has tuned at least one post-training adapter to it, not because that manufacturer has shipped the model to customers or signed off publicly. Readers should not read '17 manufacturers' as 17 endorsements.
On the benchmark front, the vendor's claims are worth deconstructing before being repeated. The marquee comparison runs on the GM-100 dual-arm suite attributed to Shanghai Jiao Tong, evaluated on the AgileX Cobot Magic and Galaxea R1 Pro: Lingbo reports LingBot-VLA 2.0 leading both π0.5 (Physical Intelligence's open generalist policy model) and NVIDIA's GR00T N1.7 on average task-progress and success rate when all three models are deployed as a single generalist, without per-task specialist fine-tuning. A second vendor-reported comparison runs the model on an Ark arm paired with an AgileX base and on the Astribot S1, and reports leads over π0.5 on long-horizon mobile-manipulation, especially in cross-domain scenes.
Those setups are deliberately apples-to-apples in one way that flatters the generalist thesis: no model in the comparison is allowed to be retuned per task. That holds the cross-embodiment promise constant, but it does not tell a buyer whether π0.5 with a task-specific fine-tune would still lose, nor whether the GM-100 numbers, labeled '初步对比测试' (preliminary comparison testing) in Lingbo's materials, reflect independent benchmarking or a benchmark Lingbo itself helped build.
The reported post-training recipe is roughly 150 demonstrations per new robot body, small enough that an integrator could port a new arm in days rather than months. The model is available on GitHub and the HuggingFace collection, alongside a technical report.
For a field that has spent three years failing to consolidate, Lingbo is now publicly betting on a concrete generalist-brain formula: a cross-embodiment base model, a low-demonstration post-training recipe, and open weights, all in one package. The bet does not say the architecture has won; it puts the hypothesis where a buyer can test it, on hardware already shipping.
Two questions remain: do the vendor-reported numbers hold up on long-horizon, real-world deployments where environmental clutter, wear, and unpredictable human behavior live; and do competing labs, from Physical Intelligence to NVIDIA, converge on a shared base-model architecture or split into separate, body-specific stacks anyway?