Robbyant Releases LingBot-VLA 2.0: Open-Source 6B Vision-Language-Action Model for Cross-Embodiment Robot Manipulation

LingBot-VLA 2.0 is an open-source 6B-parameter model released under Apache-2.0. Pretrained on 60,000 hours of data (50,000 hours of robot trajectories across 20 robot configurations and 10,000 hours of egocentric human videos), it maps all embodiments to a 55-dimensional normalized action space covering arms, dexterous hands, waist, head, and mobile base. It uses token-level, auxiliary-loss-free Mixture-of-Experts (MoE) action expert layers to scale capacity without load balancing loss. By adding geometric and temporal supervision through dual-query distillation from LingBot-Depth and DINO-Video, it achieves future-oriented control. On the GM-100 benchmark, it outperforms π0.5 and LingBot-VLA-1.0 on both evaluation platforms. For developers working on cross-embodiment robot manipulation, LingBot-VLA 2.0 offers a unified action space and efficient MoE architecture, significantly reducing the cost of transferring policies across different robot platforms. Its Apache-2.0 license also facilitates academic and commercial use.