Daily Digest — 2026-07-12

Ant Group’s Robbyant Unveils LingBot-VA 2.0: A Causal Video-Action Model Built Natively for Physical AI

MarkTechPost · Asif Razzaq · 2026-07-11

Robbyant introduces LingBot-VA 2.0, a causal video-action foundation model natively pretrained for embodied AI, addressing limitations of repurposed video generators. The architecture combines a semantic visual-action tokenizer (96 latent channels) with a sparse MoE-based causal DiT (15.3B parameters, ~2.5B active/token), employing multi-chunk prediction and foresight reasoning for asynchronous control. Evaluated on RoboTwin 2.0, it achieves 93.6% average task success, with optimizations reducing chunk latency from 927ms to 142ms (225Hz async control). Key innovations include aligned state-action latents and hybrid training objectives (T2I, T2V, TI2VA).

causal ditvisual-action tokenizersparse moeforesight reasoningmulti-chunk prediction

Read original →