Tournament era for open robotics: LingBot-VLA 2.0 becomes OpenRoboto's new base model
Robbyant's Apache-2.0 vision-language-action model is now the base of a community tournament on OpenRoboto: everyone fine-tunes, everyone is measured under identical conditions, and only a strictly better model becomes the next base.
A striking collaboration has landed on the open-source robotics front: LingBot-VLA 2.0, Robbyant's open-weight vision-language-action (VLA) model, has been declared the new base model on the community platform OpenRoboto. The 6-billion-parameter model ships under Apache 2.0, and the launch came with mutual applause; OpenRoboto thanked the Robbyant team "for releasing work that most labs would have kept to themselves".
The dataset behind it is among the field's broadest: 50,000 hours of real robot trajectories across 20 robot configurations, plus 10,000 hours of egocentric human video. The real claim is generality: a single unified action space drives arms, grippers, dexterous hands, waists, heads and mobile bases. The 20 supported embodiments include familiar names like Unitree G1, Fourier GR-2, Franka and AgiBot, and the team put the method on arXiv ("From Foundation to Application: Improving VLA Models in Practice").
The freshest idea here is not the license but the mechanism: OpenRoboto runs the whole thing as a tournament. Developers pull the model, fine-tune it, get measured under identical conditions, and publish results openly; only a strictly better model becomes the next base. Every gain carries forward into the trunk instead of sitting forgotten in a fork. An open model, they say, is only as good as what people do with it next; this design turns that sentence into a system.
The bigger picture: this week we covered Kimi K3 moving onto rentable infrastructure; the open-evolution playbook that works for language models is now being assembled for robotics. Open data, open weights, shared measurement; closed robotics labs are getting a distributed competitor.
