Decentralized Real-Time Planning for Multi-UAV Cooperative Manipulation via Imitation Learning

TL;DR

This work proposes a novel machine learning based method for decentralized kinodynamic planning that operates effectively under partial observability and without inter-agent communication for transporting and manipulating cable-suspended loads using multiple UAVs along reference trajectories.

Abstract

Existing approaches for transporting and manipulating cable-suspended loads using multiple UAVs along reference trajectories typically rely on either centralized control architectures or reliable inter-agent communication. In this work, we propose a novel machine learning based method for decentralized kinodynamic planning that operates effectively under partial observability and without inter-agent communication. Our method leverages imitation learning to train a decentralized student policy for each UAV by imitating a centralized kinodynamic motion planner with access to privileged global observations. The student policy generates smooth trajectories using physics-informed neural networks that respect the derivative relationships in motion. During training, the student policies utilize the full trajectory generated by the teacher policy, leading to improved sample efficiency. Moreover, each student policy can be trained in under two hours on a standard laptop. We validate our method in both simulation and real-world environments to follow an agile reference trajectory, demonstrating performance comparable to that of centralized approaches.

Figures (16)

• Figure 1: We enable decentralized cooperative aerial manipulation through student policies that operate independently using only the ego UAV’s state and the pose of the load. These student policies are trained via imitation learning from a centralized teacher policy with privileged observations, including the full state of the other UAVs and the load. The policy has been tested in real-world environments, where three UAVs cooperatively manipulate a cable-suspended load.
• Figure 2: The architecture of the student policy is made up of two parts - Encoder (Sec. \ref{['sec:obs_space']}) & Decoder (Sec. \ref{['sec:action_space']}). The Encoder network maps the observation histories to a latent vector $x_t$. This latent vector is then copied for all nodes in the horizon and passed to the Decoder network.
• Figure 3: Top view of tracking performance. Circle, cross indicate start, end of trajectory. Two laps of Eight are performed.
• Figure 4: Despite lacking awareness of each other's positions, the quadrotors successfully maintain safe inter-agent distances
• Figure 5: Left: Performance of the system in simulation | Right: Performance of the system when deployed in the real world | Bottom: Performance of the second version of student policies in real world.