Ant Group's Robbyant Releases LingBot-VA 2.0 Embodied AI Model

Robbyant, the embodied AI division of Ant Group, has released LingBot-VA 2.0, an embodied-native foundation model specifically designed for generalist robot manipulation. This new model represents a significant advancement by pretraining the entire stack for embodiment, rather than relying on fine-tuning existing video generation components. LingBot-VA 2.0 addresses limitations found in previous video-action models, which often reuse elements built for digital content creation. These older models typically employ a reconstruction-oriented VAE and a bidirectional video-diffusion backbone with an attached action module. This approach results in pixel-reconstruction latents that preserve appearance but offer limited physical structure, and iterative denoising processes that are too slow for real-time closed-loop control. Furthermore, generic video objectives do not adequately teach how actions physically alter the environment, and the use of bidirectional attention in backbones conflicts with the strictly forward progression of control in time.

LingBot-VA 2.0's architecture diverges from these predecessors. Version 1.0 of the model already began adapting the existing stack into a causal model. However, version 2.0 takes this a step further by natively pretraining a causal DiT (Diffusion Transformer). The first stage of LingBot-VA 2.0 replaces the traditional compression-only VAE with a new tokenizer. This tokenizer incorporates two additional objectives beyond simple reconstruction: semantic alignment, which aligns visual latents with a frozen Perception Encoder teacher, and a latent-action objective that extracts compact transition variables between consecutive latents. An inverse dynamics model is used to predict each latent action, which is then decoded by a forward dynamics model into a transport map and a residual. This integration allows world states and actions to share a single latent space, enabling unlabeled web video to provide action-relevant supervision.

The second stage of LingBot-VA 2.0 features a causal DiT with a sparse Mixture-of-Experts (MoE) video stream. This design maintains the MoE layout from version 1.0, where a video expert and an action expert share a causal self-attention mechanism. Each expert has its own dedicated feed-forward pathway, allowing for specialized processing while maintaining temporal causality crucial for robot control. The model's GitHub repository provides access to the research paper detailing these advancements.