Archon Robotics Presents Whole-Body AI Control Across Humanoids

SHANGHAI -- Archon Robotics unveiled a whole-body AI control system designed to coordinate walking, balancing and object manipulation across different humanoid robots, showing the technology in an Oct. 12 demonstration involving machines from several manufacturers.

The Shanghai-based company calls its approach Whole-Body Intelligence. It combines training on human movement data, while extra training helps robots adapt, aiming to develop AI policies that control a robot's entire body rather than treating locomotion and manipulation as separate problems.

In its debut demonstration, Archon showcased on humanoid platforms from AGIBOT, Booster Robotics and a modified Unitree. The company says its approach can be modified for various computing architectures.

The demonstration does not establish how consistently the system performs across those platforms. Archon has not published comparative benchmarks, task success rates, or human-intervention counts.

Humanoid robots typically require hierarchical control systems. Walking involves balancing and foot placement, while manipulation requires arm-and-hand control in relation to objects and the environment.

The tasks become more difficult when a robot must reach for, carry, or even manipulate objects while moving. Changes in arm position or the weight of an object can affect the robot's balance. Archon's approach aims to coordinate movements under one shared control framework.

The company describes a training process that starts with the capture of human movement data and continues with robot-specific training. Human data can provide examples of coordinated motion, while extra training helps robots adapt for differences in joint configurations, actuators and physical capabilities.

A system capable of transferring learned behaviors between robot platforms could eliminate the need to train separate policies for each robot. However, Archon hasn't revealed sufficient technical detail.
The use of several humanoid platforms is a central part of Archon's presentation. Robot manufacturers differ in body proportions, joint arrangements and actuator capabilities, making it hard to transfer learned movement policies.

The demonstration suggests Archon is targeting a software layer capable of working across different robot designs, not just a single proprietary humanoid. No paying customers or deployments have been announced.

It also has not released the model weights, training dataset sizes or standardized evaluation results needed to compare its approach with other humanoid control systems.