Inside the LingBot-VA 2.0 Robotic Foundation Model

Aiden_Tech_Ai · x · 2026-07-11

This post breaks down the LingBot-VA 2.0 paper, arguing that the foundation model for robotic control is more critical than downstream fine-tuning.

Core Design

Rather than hacking a video generation model, it pre-trains a native video-action foundation model from scratch specifically for robot control.
It employs a shared semantic vision-action tokenizer to align world states and latent actions into the same semantic space.
It leverages causal pre-training from the very beginning to avoid the catastrophic forgetting that occurs when adapting bidirectional video models into control models.
The architecture integrates a roughly 13B sparse MoE video backbone, but only activates about 1.9B parameters per token to balance capacity with high-frequency control efficiency.
It also introduces Foresight Reasoning, enabling the model to predict and continuously re-align before executing actions.

More from Embodied