RoboScience, a Chinese robotics company, has launched Visics-Agent, a cloud-based embodied AI service that it says allows robots to understand voice commands, remember user habits and learn new skills.
The service, offered as an embodied AI service (EaaS), is aimed at robot developers. Developers can connect dexterous hands to its API through Tencent Cloud without collecting data, writing code or fine-tuning models, according to the company. Users can issue instructions directly and the robot will complete tasks autonomously.
Visics-Agent separates robot learning into two stages: teaching and execution. In the teaching stage, users can instruct a robot in natural language, correct it if needed and tell it to remember a successful result. The system then turns the experience into a reusable skill that can be called up for similar tasks and combined with other skills in new scenarios.
RoboScience has also designed a Harness system that uses real-world results and feedback to refine existing skills. In one desktop-cleaning task, the robot learned to lift its arm to clear the camera view before planning its next move. After a month of continuous real-robot testing, Visics-Agent accumulated dozens of effective experiences and raised the success rate of a single skill to more than 85%, the company said.
In internal tests covering 454 task instructions, success rates for L1 and L2 tasks, which involve one or two pick-and-place operations, exceeded 90%, while complex L3 tasks requiring three to seven continuous operations reached 78%, according to RoboScience.
Visics-Agent is powered by RoboScience’s Visics general manipulation model and supports different end effectors, including dexterous hands and two- or three-finger grippers, with switching in under 30 seconds, the company said. The company demonstrated the system driving a desktop robotic arm at Tencent’s Global Digital Ecosystem Summit.
Visics-Agent is scheduled to launch in December 2026. First-batch applications are open to robot developers, researchers and companies exploring robot deployment, RoboScience said.
AGIBOT
Figure AI
UBTECH