AI News

Apple research team discloses humanoid robot control system HALO, validated on Unitree G1 with an average success rate of 91%

YicaiUpdated Nov 7 19:04

AI summary

The Apple research team submitted the paper "Perceptive Humanoid Motion from Simulation" to arXiv, disclosing the autonomous humanoid robot control system HALO and deploying it on the Unitree G1, enabling motions learned in simulation to be executed in unfamiliar real rooms. The team completed 360 single-task tests across 6 real environments that had not been pre-scanned or mapped, achieving an average success rate of 91%; 3 composite tasks were each tested 10 times, with an average success rate of 70%, among which "duck under an obstacle—walk to the target—push the box" was 50%. The research team publicly released the data generation, model training, and deployment code, and noted that adding new physical tasks still requires manually writing prompts and designing motion keyframes; the robot currently relies mainly on visual perception, and incorporating tactile perception is one direction for future improvement.

Why read this

AI score 62

The paper publicly releases the data generation and deployment code for simulation-to-real-machine transfer, and provides per-scenario success rates for comparison with existing robot control approaches.

Topics

Tags

Body · AI translation

According to The Paper, Apple's research team has disclosed progress on research into a robot "brain." On November 4 local time, Apple's research team submitted a paper titled "Perceptive Humanoid Motion from Simulation" to the preprint platform arXiv, disclosing the HALO autonomous control system for humanoid robots. The research team deployed the system on Unitree's G1 humanoid robot for validation, enabling the robot to carry out motions learned in a simulated environment in unfamiliar real-world rooms. The research team also released the code for data generation, model training, and deployment, covering motion generation, motion tracking, task orchestration, and operation in both simulation and on the real robot. According to the paper, the research team conducted 360 single-task tests in 6 real environments that had not been pre-scanned or mapped, with an average success rate of 91%. In addition, in one real environment, they conducted 10 tests each for 3 combined tasks, with an average success rate of 70%. Among them, the combination of "bend down to pass an obstacle—walk toward the target—push the box" had a success rate of 50%.

It should be noted that simulation-generated data has not completely eliminated human involvement. The team said that new physical tasks still require manually written prompts and manually designed motion keyframes, and some motions also used human videos or motion-capture footage. The research team also noted that the robot currently relies mainly on visual perception and finds it difficult to directly perceive contact conditions; when an object leaves the camera's field of view or is occluded, performance on tasks involving more contact decreases. Adding tactile perception is one of the directions for future improvement.

Source

Yicai