An open-source vision-language-action system enables a single AI model to operate across multiple robot platforms without task-specific retraining.
LingBot-VLA 2.0 is an open-sourced vision-language-action (VLA) model that seeks to integrate numerous robot platforms with one AI model. Unlike most currently available robotics models which need to be retrained for each robot hardware configuration, LingBot-VLA 2.0 operates across single-arm, dual-arm, wheeled, and humanoid robots.
According to the firm, the model was trained on 60,000 hours of real-world data, including 50,000 hours of robot interaction data and 10,000 hours of first-person human manipulation data. The training set consisted of 20 robot morphologies from manufacturers of 17 different companies and helped the model achieve control of robotic arms, hands, head, waist, and mobile chassis through a common software.
On Shanghai Jiao Tong University’s GM-100 benchmark for dual-arm manipulation, Robbyant stated that their model outperformed π0.5 and GR00T N1.7 on the GM-100 benchmark. In addition, LingBot-VLA 2.0 achieved greater task progress and success rates compared to π0.5 in long-horizon mobile manipulation tasks with ARX Arm + AgileX Chassis and Astribot S1.
The company also launched a deployment-ready version of the model with an inference latency below 130 ms on an NVIDIA RTX 4090, seeking to ease hardware compatibility and post-training work in order to be applied in commercial robotics.
According to Robbyant, the model is currently being tested for use in retail sorting, logistics and industrial automation applications. The launch follows the company’s recent open-source release of LingBot-Vision and LingBot-Depth 2.0, completing the software stack for embodied AI that includes robotic perception, spatial awareness and autonomous task execution for robots.
