LingBot-VA 2.0: A Control-Native Foundation Model for Robotics

Robbyant, an embodied AI company under Ant Group, introduced LingBot-VA 2.0, a video-action foundation model. Unlike approaches that adapt general video generators with attached action heads, this model's entire stack is natively pre-trained for robotic control from the ground up. This architectural choice has sparked discussions on reducing demonstration costs and improving control efficiency.

Key Design and Architecture

According to rohanpaul_ai, LingBot-VA 2.0's control policy is trained causally from the start rather than retrofitting a bidirectional video model. Its video stream utilizes 128 routed experts, top-8 routing, and 1 shared expert. For representation learning, it moves beyond pixel-reconstruction tokenizers by aligning compressed states with semantic features from a frozen vision model, incorporating forward and inverse dynamics. To prevent the model from simply copying adjacent frames, it employs Multi-Chunk Prediction with 3 lightweight heads to predict the next 1, 2, and 3 chunks.

Inference Efficiency and Results

System optimizations include distilling inference down to 2 function evaluations per chunk. Combined with FP8 TensorRT, it achieves a 6.5x speedup and a 225Hz control frequency. On the RoboTwin 2.0 benchmark, it reports an average score of 93.6 (93.8 for clean, 93.4 for randomized).

External Interpretations

Div_pradeep echoed the sentiment that being "control-natively pre-trained" is more critical than sheer parameter count. annbordetsky highlighted the system's goal of reducing the expensive demonstration costs typically required when a robot's body, task, or environment changes.

2026-07-12 ~ 2026-07-14 · 9 related posts