# ARMOR: Egocentric Perception for Humanoid Robot Collision Avoidance and Motion Planning

**Authors:** Daehwa Kim, Mario Srouji, Chen Chen, Jian Zhang  
**arXiv:** [2412.00396](https://arxiv.org/abs/2412.00396)  
**Date:** 2024-11-30  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2412.00396v1/

## Key points

- Challenges faced by humanoid robots in sensing and perception while navigating dense environments
- Development of ARMOR, an egocentric perception system for enhancing collision avoidance and motion planning
- Integration of hardware and software components, including wearable-like depth sensors tailored for humanoid robots
- Key feature: distributed perception approach improving spatial awareness and enabling agile motion planning capabilities
- Implementation of a transformer-based imitation learning (IL) policy trained on human realistic motions from the AMASS dataset
- Results showing significant improvements over traditional setups, with a 63.7% reduction in collisions and a 78.7% improvement in success rate
- Comparison with cuRobo demonstrating superior performance of IL policy with a 31.6% decrease in collisions, 16.9% higher success rate, and 26x reduction in computational latency
- Deployment of ARMOR on a GR1 humanoid robot for real-world validation
- Future plans to update publication with source code, hardware descriptions, and CAD files for further research and development

## Summary

In the field of robotics, humanoid robots face challenges in sensing and perception while navigating through dense environments. To address this issue, a team of researchers led by Daehwa Kim, Mario Srouji, Chen Chen, and Jian Zhang have developed ARMOR - an innovative egocentric perception system designed to enhance collision avoidance and motion planning for humanoid robots. This system integrates both hardware and software components, including wearable-like depth sensors specifically tailored for humanoid robots. The key feature of ARMOR is its distributed perception approach which significantly improves the robot's spatial awareness and enables more agile motion planning capabilities. Additionally, the researchers implemented a transformer-based imitation learning (IL) policy in simulation as part of the hardware setup. This IL policy was trained using approximately 86 hours of human realistic motions extracted from the AMASS dataset. The results demonstrate the effectiveness of ARMOR compared to traditional setups utilizing multiple dense head-mounted and externally mounted depth cameras. The ARMOR perception system achieved a remarkable 63.7% reduction in collisions and a substantial 78.7% improvement in success rate. Furthermore, the researchers compared their IL policy against a sampling-based motion planning expert cuRobo. The results showed that the IL policy outperformed cuRobo with a 31.6% decrease in collisions, a 16.9% higher success rate, and an impressive 26x reduction in computational latency. To validate their findings in real-world scenarios, the team deployed the ARMOR perception system on a GR1 humanoid robot from Fourier Intelligence. Moving forward, they plan to update the arXiv version of their publication with links to source code, hardware descriptions, and 3D CAD files for further research and development. Overall,in enhancing humanoid robot capabilities through innovative egocentric perception systems like ARMOR. By addressing gaps in sensing and perception, these advancements pave the way for more efficient and reliable navigation in complex environments for humanoid robots.