LingBot-VA 2.0: A 225Hz Robot Control Model

rohanpaul_ai · x · 2026-07-14

Here is a comprehensive overview of LingBot-VA 2.0, highlighting the core aspects:

What it is

A video-action foundation model developed by Robbyant (an embodied AI company under Ant Group), designed to enable robots to directly "see, predict, and control" in the real world.
The authors emphasize it is not merely a video generation model with an action head attached, but is designed from the ground up for control tasks.

Key methods

Employs a semantic visual-action tokenizer: compresses observations into a semantic space closely aligned with a frozen vision model, and learns latent actions from inter-frame changes.
Utilizes self-supervised inverse/forward dynamics, allowing unlabeled web videos to provide action-related training signals.
The policy is inherently causal, making decisions based solely on past observations.
The video backbone uses MoE: approximately 13B total parameters with around 1.9B activated per token.
A high-level VLM planner decomposes long-horizon tasks, while a low-level video-action policy handles continuous control.

Execution and acceleration

Foresight Reasoning predicts future visual states while the robot executes its current action, correcting the cache with real frames whenever new observations arrive.
Combined with few-step distillation and system optimizations, the paper reports a peak asynchronous execution frequency of 225 Hz.

Results

Capable of adapting to tasks from just 10–15 demonstrations.
Supports cross-embodiment transfer and demonstrates zero-shot capabilities on certain new tasks.
In the paper's own evaluations, it achieved an average score of 93.6 on RoboTwin 2.0, outperforming both LingBot-VA and π0.5 on real-world tasks.

More from Embodied