Ant Group's LingBot-VA 2.0: a robot brain that runs on one GPU

Most frontier models are trained on text and images and reason about the world. A world action model is trained to output actions -- motor commands and manipulation plans -- based on what a robot sees, so it can directly control hardware. "Embodied-native" means built for that from the ground up, not a language model bolted onto a robot arm. Ant says LingBot-VA 2.0 is the first of its kind, a framing that is itself part of the pitch rather than an independently settled fact.

In demos the model handles delicate objects -- holding potato chips without crushing them -- and performs everyday manipulation such as tidying a desk, the kind of soft, force-sensitive tasks that are notoriously hard for robots. The two figures Ant leads with: a 93.6% success rate in simulation tests, and the claim that it runs on a single GPU, positioning it for low-cost, wide deployment. That efficiency angle is the real story, because it targets the cost that usually keeps capable robot control out of everyday machines. The launch lands in a busy week for embodied AI, alongside Rhoda AI's FutureVision and Mecka AI's robot-action-data business, continuing China's efficiency-and-deployment focus.

Watch for independent, real-hardware benchmarks -- not simulation -- and whether the single-GPU claim holds under real robot workloads with unpredictable objects and lighting. If it does, cheap on-device robot control is the genuinely disruptive part, because it lowers the cost of putting capable manipulation on ordinary machines. If it doesn't, this stays a promising demo among many that landed this same week.