LingBot-VA2.0: Embodied Native VA Model

量子位 · wechat · 2026-07-10

Ant Lingbo (蚂蚁灵波) released LingBot-VA2.0, the world's first embodied native pre-trained VA foundation model. It enables robots to predict actions based on video and plan ahead rather than merely reacting. The article demonstrates real-world robotic tasks like desk organizing, conveyor belt grasping, and picking up potato chips, highlighting its long-term memory, temporal alignment, and fine-grained manipulation capabilities.

The technical approach includes a novel semantic vision-action tokenizer and a causal DiT combined with a sparse MoE backbone. Results show a 93.6% success rate in bimanual tasks and 150Hz inference on a single GPU. Overall, it illustrates how embodied native VA models enhance predictive control for robots in complex physical environments.

Related event: LingBot-VA/VLA 2.0 Released: Native Embodied Foundation Model(24 posts)→

Original post →

More from Embodied

Embodied channel →