DMCA: Dense Multi-agent Navigation using Attention and Communication

TL;DR

This work uses a multi-head self-attention mechanism that encodes observable information from neighboring robots into a concise and fixed-length observation vector, thereby handling varying numbers of neighbors and focuses on improving navigation performance through selective communication.

Abstract

In decentralized multi-robot navigation, ensuring safe and efficient movement with limited environmental awareness remains a challenge. While robots traditionally navigate based on local observations, this approach falters in complex environments. A possible solution is to enhance understanding of the world through inter-agent communication, but mere information broadcasting falls short in efficiency. In this work, we address this problem by simultaneously learning decentralized multi-robot collision avoidance and selective inter-agent communication. We use a multi-head self-attention mechanism that encodes observable information from neighboring robots into a concise and fixed-length observation vector, thereby handling varying numbers of neighbors. Our method focuses on improving navigation performance through selective communication. We cast the communication selection as a link prediction problem, where the network determines the necessity of establishing a communication link with a specific neighbor based on the observable state information. The communicated information enhances the neighbor's observation and aids in selecting an appropriate navigation plan. By training the network end-to-end, we concurrently learn the optimal weights for the observation encoder, communication selection, and navigation components. We showcase the benefits of our approach by achieving safe and efficient navigation among multiple robots, even in dense and challenging environments. Comparative evaluations against various learning-based and model-based baselines demonstrate our superior navigation performance, resulting in an impressive improvement of up to 24% in success rate within complex evaluation scenarios.

Figures (20)

• Figure 1: We illustrate the high-level network architecture used for multi-agent navigation in DMCA. Primarily, the network consists of three modules: the observation encoder, the communication selection, and the navigation block.
• Figure 2: We compare the trajectories generated by our proposed method with CADRL and GA3C-CADRL for a circular scenario with 20 agents. We observe that DMCA generates smooth and collision-free trajectories to the goal, while CADRL results in some collisions. In GA3C-CADRL, some agents were deadlocked while others were in a collision, and no agent reached the goal. The network requires 6 ms per agent to compute an action.
• Figure 3: We consider a complex scenario where the robots are arranged in an $n \times n$ grid formation. The final formation is created by moving the robot at position $(\texttt{row},\texttt{column})$ to $(\texttt{n-row+1}, \texttt{n-column+1})$. Figure \ref{['init']} shows the initial configuration of the agents, Figure \ref{['NavSIC']}- \ref{['RVO']} shows the final configuration for DMCA, DMCA-LC, CADRL, and ORCA. We observe that DMCA generates collision-free trajectories in this scenario while DMCA-LC results in one robot being deadlocked. CADRL results in some robots colliding, while results in agents colliding or deadlocked. Thus DMCA and DMCA-LC perform the best in this scenario.
• Figure 4: We consider a swap scenario with disk-shaped static obstacles (black) and eight agents. We observe that the eight agents using DMCA remain safe and successfully reach their goal in this scenario.
• Figure 5: Real-world Scenario: We evaluate DMCA in a real-world setting with three agents in the circle scenario. Fig 5 (a)-(e) present the snapshot of the execution every 10 seconds. Fig 5(f) illustrates the trajectories followed by each agent. We observe the robots safely navigate to their goal position in this scenario.