LingBot-VA 2.0 launches robot control model
Developed by Robbyant under Ant Group, LingBot-VA 2.0 is a video-action foundation model built from scratch for native robot control. It employs a causal Mixture-of-Experts architecture and consistency distillation to reduce control loop latency to 142 ms.
By pretraining a causal video-action stack from scratch instead of adapting bidirectional models, LingBot-VA 2.0 preserves physical dynamics and control priors.
* **Causal over Bidirectional:** Adapting bidirectional generators for control loses temporal priors; causal training keeps the model aligned with forward-moving control loops.
* **Scaling via Unlabeled Video:** Self-supervised action learning from web video mitigates the data scarcity bottleneck of physical robot demonstrations.
* **System-Modeling Co-design:** Distillation and hardware-level optimization (FP8 TensorRT) achieve the low latency necessary for practical, closed-loop deployment.
* **Sparse MoE Efficiency:** The 13B MoE architecture scales capacity for complex visual dynamics while maintaining a low active parameter count (1.9B) per token.
DISCOVERED
1h ago
2026-07-11
PUBLISHED
1h ago
2026-07-11
RELEVANCE
AUTHOR
omarsar0
