# Spatial Intention Maps for Multi-Agent Mobile Manipulation

**Authors:** Jimmy Wu, Xingyuan Sun, Andy Zeng, Shuran Song, Szymon Rusinkiewicz, Thomas Funkhouser  
**arXiv:** [2103.12710](https://arxiv.org/abs/2103.12710)  
**Date:** 2021-03-23  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2103.12710v1/

## Key points

- "Spatial Intention Maps for Multi-Agent Mobile Manipulation" study conducted by Jimmy Wu, Xingyuan Sun, Andy Zeng, Shuran Song, Szymon Rusinkiewicz, and Thomas Funkhouser
- Focuses on enhancing coordination and collaboration among decentralized multi-agent robots during physical tasks using spatial intention maps
- Spatial intention maps are a novel representation for vision-based deep reinforcement learning in multi-agent scenarios
- Each agent's intention is communicated to other agents and translated into an overhead 2D map aligned with visual observations
- Integration complements the spatial action maps framework to foster emergent cooperative behaviors like object passing or collision avoidance
- Experiments conducted across various multi-agent environments with heterogeneous robots show significant performance improvements with spatial intention maps
- Enhances efficiency of mobile manipulation tasks and boosts cooperative behaviors among decentralized robot team members
- Research to be presented at IEEE International Conference on Robotics and Automation (ICRA) in 2021 showcasing potential for advancing collaborative robotics through enhanced communication and coordination mechanisms via spatial intention mapping

## Summary

"Spatial Intention Maps for Multi-Agent Mobile Manipulation" is a groundbreaking study conducted by Jimmy Wu, Xingyuan Sun, Andy Zeng, Shuran Song, Szymon Rusinkiewicz, and Thomas Funkhouser. The research focuses on enhancing coordination and collaboration among decentralized multi-agent robots during physical tasks through the utilization of spatial intention maps. These maps serve as a novel representation for vision-based deep reinforcement learning in multi-agent scenarios. In this innovative approach, each agent's intention is communicated to other agents and translated into an overhead 2D map aligned with visual observations. This integration complements the spatial action maps framework and aligns state and action representations spatially to foster emergent cooperative behaviors such as object passing or collision avoidance. Experiments were conducted across various multi-agent environments with teams of heterogeneous robots possessing diverse capabilities like lifting, pushing, or throwing objects. The results demonstrate significant performance improvements achieved by incorporating spatial intention maps. Not only do these maps enhance the efficiency of mobile manipulation tasks but they also notably boost cooperative behaviors among decentralized robot team members. This research will be presented at the IEEE International Conference on Robotics and Automation (ICRA) in 2021 and showcases a promising avenue for advancing collaborative robotics through enhanced communication and coordination mechanisms facilitated by spatial intention mapping. Further insights into this cutting-edge work aimed at revolutionizing multi-agent mobile manipulation systems can be found on the project page.