Xenon to Build AI That Lets Humanoid Robots Fix Their Own Errors

Fixing the break in multi-step tasks

The goal is to let a humanoid notice and respond to errors while carrying out tasks made of several steps. According to the company, current robots can succeed at individual actions such as grasping an object or opening a door, yet the whole task can stall if the robot fails to reach a suitable position or posture for the next step.

Xenon plans an action-supervision system that analyzes the robot's state, its task plan and past successes and failures. It will be combined with a world model that predicts future situations from the results of actions. Together they are meant to check progress in real time and, when something goes wrong, revise instructions or propose a new path.

The company's example: a robot opens a box, takes out a can and puts it in a refrigerator. If it reaches a point where the next motion is hard to perform, the system adjusts the instruction or suggests a recovery method. The aim is less human intervention and longer continuous operation.

Path to commercialization

Xenon opened a Physical AI Lab in March and is building an AI engine for humanoid perception, judgment and action, along with a robot operations platform called Physical Ops. It will also reuse technology from its enterprise generative AI platform GenOS.

Results will be applied to the company's own humanoid, GenP, and tested in real work settings. Xenon then plans to expand into welfare-facility living support and docent services, and to commercialize through licenses and subscriptions for robot operation and management software.

Chief Technology Officer Myung Dae-woo said the selection is an important chance to extend the company's enterprise AI strengths into physical AI, and that Xenon will work to turn the results into real services.