Ant Group Open-Sources LingBot-VLA 2.0

Robbyant, an Embodied AI company under Ant Group, has open-sourced LingBot-VLA 2.0. The core of this model is not a complex architectural change but rather the choice of representation using Relative Joint Actions, which significantly increased the success rate. Experimental results showed that simply switching the policy from absolute targets to relative targets raised the average success rate from 33.7% to 55.0%, a 21.3%p increase.

Key technical specifications:

• Unified action vector: Uses a 55-dimensional canonical action vector covering arm joints, end effectors, grippers, 12-dimensional dexterous hands, waist, head, and mobile base
• Generality: Trains a single policy for 20 robot models, ranging from 8-DoF single arms to 32-DoF humanoids
• Dataset: Pretrained on approximately 60,000 hours of data (50,000 hours of robot trajectories + 10,000 hours of first-person human videos)
• Model structure: Applies a DeepSeek-V3-style MoE (Mixture of Experts) action expert along with Dual-query distillation technology

Both the model weights and code are publicly available, allowing direct testing across various robot environments.