World’s First Cloud‑Deployed Embodied AI Model Swaps Robotic Hands in 30 Seconds

At WAIC 2026, Visics unveiled the first cloud‑deployed embodied large model, proving that a single AI "brain" can control multiple robotic hands and switch between them in just 30 seconds without any on‑site data collection or retraining.

The demo began with a freshly grasped object using one dexterous hand, which was then physically removed and replaced by a different hand. After a 30‑second pause, the robot resumed operation, handling new objects supplied by the audience. This "one brain, many hands" capability was repeated across three challenge rounds, each involving a different hand, object, and task.

Swapping end‑effectors traditionally requires a full retraining cycle because each hand changes the action space, control dynamics, and hardware parameters—effectively a new "employee" that must be onboarded. Visics’ VLOA (Vision‑Language‑Object‑Action) model eliminates this training bottleneck by leveraging zero‑shot generalization: the same model instantly adapts to over ten different hand brands without additional data collection.

The VLOA architecture consists of an upper world model that predicts the 3‑D object trajectory (the "Object Trajectory" token) and a lower operation model that translates this trajectory into concrete robot commands—contact point, approach direction, and force. A self‑developed FingerEye module provides depth guidance before contact and six‑axis force/torque feedback after contact, enabling sub‑millimetre force control for delicate tasks such as opening red envelopes and stacking coins.

Visics trains the model on a massive dataset of video and simulated manipulation sequences. Their automated labeling pipeline reduces data acquisition cost to 1/20–1/200 of traditional methods, and the pipeline has already generated hundreds of billions of manipulation trajectories, with a target of reaching a trillion‑scale dataset by 2026.

Building on this foundation, Visics partnered with Tencent Cloud to launch Embodied AI as a Service (EaaS). The model, its data standards, and the Object Trajectory token are exposed via APIs, allowing developers to invoke the same intelligence across heterogeneous hardware without re‑engineering the robot’s software stack.

This cloud‑first approach decouples the robot’s physical body from its intelligence, turning embodied AI from a bespoke, hardware‑bound solution into a scalable, updatable service. While latency, security, and network reliability remain challenges, the demonstration signals a clear shift toward standardized, platform‑level embodied AI for a wide range of industries.