Show HN: Open-weights VLA model for 20 robot embodiments (code and checkpoints) (github.com)

LingBot-VLA 2.0 has been announced as an advanced Vision-Language-Action foundation model aimed at enhancing real-world robotic applications. It builds on its predecessor, LingBot-VLA 1.0, by significantly improving generalization across various tasks and robot configurations, expanding the range of actions robots can perform, and integrating predictive dynamics modeling. With over 60,000 hours of training data, including diverse robot trajectories and egocentric human videos, this update provides a unified representation for various robotic embodiments, from simple arms to dexterous hands. The model employs a sophisticated architecture that incorporates sparse Mixture of Experts (MoE) layers and uses novel mechanisms for future scene prediction, thereby encouraging causal inference. By mapping multiple robot embodiments into a comprehensive 55-dimensional state/action vector, it supports more dynamic interactions. These enhancements are particularly pivotal as they allow for more flexible and robust robotic behaviors in complex environments, paving the way for improvements in efficiency and effectiveness across varied manipulation tasks. Researchers and developers can access pre-trained weights and detailed configuration guides, making LingBot-VLA 2.0 a valuable resource for advancing AI-driven robotics.