PRIMER: Perception-Aware Robust Learning-based Multiagent Trajectory Planner

TL;DR

The paper tackles perception-aware, decentralized multiagent trajectory planning under localization uncertainty by introducing optimization-based PARM/PARM* planners and a learning-based counterpart, PRIMER. PRIMER is trained via imitation learning using PARM* as the expert and employs an MLP with an LSTM backbone to handle variable obstacle counts, enabling fast, scalable replanning while still respecting deconfliction through the Robust MADER framework. Empirical results show PRIMER achieving up to approximately 5.6k× faster replanning with minimal loss in trajectory quality and no dynamic-constraint violations, across single- and multi-agent scenarios. This contributes a practical pathway to scalable swarm coordination in dynamic, uncertain environments, with future work focusing on larger-scale and hardware validations.

Abstract

In decentralized multiagent trajectory planners, agents need to communicate and exchange their positions to generate collision-free trajectories. However, due to localization errors/uncertainties, trajectory deconfliction can fail even if trajectories are perfectly shared between agents. To address this issue, we first present PARM and PARM*, perception-aware, decentralized, asynchronous multiagent trajectory planners that enable a team of agents to navigate uncertain environments while deconflicting trajectories and avoiding obstacles using perception information. PARM* differs from PARM as it is less conservative, using more computation to find closer-to-optimal solutions. While these methods achieve state-of-the-art performance, they suffer from high computational costs as they need to solve large optimization problems onboard, making it difficult for agents to replan at high rates. To overcome this challenge, we present our second key contribution, PRIMER, a learning-based planner trained with imitation learning (IL) using PARM* as the expert demonstrator. PRIMER leverages the low computational requirements at deployment of neural networks and achieves a computation speed up to 5500 times faster than optimization-based approaches.

Figures (5)

• Figure 1: PARM 's trajectory optimization and deconfliction sequence: PARM uses an optimization-based approach to generate trajectories for each agent, followed by a conflict detection and resolution step based on the Robust MADER framework. Each agent first generates a new trajectory in the optimization step, and then checks if there are any conflicts with the trajectories received from other agents. If no conflicts are detected, the agent publishes its new trajectory and begins checking for potential collisions in a delay check step. This delay check step is a sequence of checks over a period of time. Finally, if no conflicts are detected during the delay check, the agent commits to the new trajectory and publishes it. However, if conflicts are detected, the agent reverts to the trajectory from the previous iteration and discards the new trajectory. More details on the RMADER approach can be found in Section II of kondo2023robust_ral.
• Figure 2: PRIMER Network Architectures
• Figure 3: PRIMER Planning and Sharing Trajectory Architecture
• Figure 4: Student multi-agent, multi-obstacle, simulation result: We made three imitation learning-based (student) agents fly around two dynamic obstacles. They started at the top-right corner and was commanded to fly to the down-left. For simplicity, we omitted FOV tripods visualization.
• Figure 5: Results of flight simulations. (a) The student's computation time is much faster than that of the expert, and (b) the student's travel time is also much shorter; this is mainly because of the faster computation time. (c-d) Since the student achieves faster replanning, it does not need to stop as the expert does, and that leads to smoother trajectory generation.