Summary
Robbyant (Ant Group’s embodied AI unit) released LingBot-VLA 2.0 on July 8, 2026 as an open-source 6B-parameter VLA trained on 60,000 hours of real-world robot data spanning 20 morphologies and 17 manufacturers. The model achieves comprehensive breakthroughs in full-body DoF control, spatial perception, post-training efficiency, and long-horizon task execution with sub-130 ms inference latency.
Key Contributions
- Trained on 50,000 hours of cleaned real-robot interaction data plus 10,000 hours of distilled first-person human manipulation data across single-arm, dual-arm, bipedal, and wheeled configurations
- Extended operational coverage to head, waist, end-effectors (hands), and mobile chassis, enabling coordinated whole-body control across 20+ robot configurations from 17 manufacturers (Unitree, AgiBot, Franka, etc.)
- Sub-130 ms inference latency with strong long-horizon task execution via improved post-training efficiency recipes
Significance
One of the largest publicly available cross-embodiment VLAs, LingBot-VLA 2.0 represents a major open-source counterpart to proprietary cross-embodiment foundation models and demonstrates that diverse manufacturer collaboration can underpin truly generalist robot brains.