Auto-Multilift: Distributed Learning and Control for Cooperative Load Transportation With Quadrotors

TL;DR

Auto-Multilift presents a distributed, closed-loop framework that automatically tunes MPC hyperparameters for cooperative cable-suspended load transport with quadrotors. It models MPC costs with deep neural networks and trains them via a distributed policy gradient using closed-loop trajectories, facilitated by distributed sensitivity propagation that exploits multi-agent couplings. The method demonstrates scalability to multiple quadrotors and outperforms open-loop MPC tuning, including in obstacle-rich scenarios requiring adaptive tension references. The combination of DSP and Safe-PDP-based gradient computation enables efficient end-to-end learning directly from system tracking errors, with potential for real-time adaptation and expanded applicability to complex multi-robot transport tasks.

Abstract

Designing motion control and planning algorithms for multilift systems remains challenging due to the complexities of dynamics, collision avoidance, actuator limits, and scalability. Existing methods that use optimization and distributed techniques effectively address these constraints and scalability issues. However, they often require substantial manual tuning, leading to suboptimal performance. This paper proposes Auto-Multilift, a novel framework that automates the tuning of model predictive controllers (MPCs) for multilift systems. We model the MPC cost functions with deep neural networks (DNNs), enabling fast online adaptation to various scenarios. We develop a distributed policy gradient algorithm to train these DNNs efficiently in a closed-loop manner. Central to our algorithm is distributed sensitivity propagation, which is built on fully exploiting the unique dynamic couplings within the multilift system. It parallelizes gradient computation across quadrotors and focuses on actual system state sensitivities relative to key MPC parameters. Extensive simulations demonstrate favorable scalability to a large number of quadrotors. Our method outperforms a state-of-the-art open-loop MPC tuning approach by effectively learning adaptive MPCs from trajectory tracking errors. It also excels in learning an adaptive reference for reconfiguring the system when traversing multiple narrow slots.

Figures (13)

• Figure 1: Illustration of a multilift system and Auto-Multilift learning pipelines. (a) Components and structure of a typical multilift system; (b) A block diagram of the proposed framework. We fuse DNNs with distributed MPC controllers to obtain adaptive cost function parameters online. These DNNs can be trained efficiently in a distributed and close-loop manner. This is achieved through distributed sensitivity propagation, a key module in our method that computes actual state sensitivities in parallel across the quadrotors.
• Figure 2: Illustration of the multilift system. Let $\mathcal{I}$, $\mathcal{B}^l$, and $\mathcal{B}^i$ denote the world frame, the body frame attached to the load, and the body frame attached to the $i$-th quadrotor, respectively.
• Figure 3: Illustration of the data exchange used in the distributed MPC.
• Figure 4: Diagram of data sharing used in the DSP algorithm. As discussed in Section \ref{['subsec:distributed mpc']}, the computation for the load is completed by a randomly selected 'central' agent; therefore, the data are shared between the central quadrotor and the remaining quadrotors.
• Figure 5: Architecture of the neural network for producing the adaptive MPC normalized hyperparameters $\boldsymbol{\Theta}^i$ online. As the network's input, the observations can include the system tracking errors or obstacle information, depending on the applications.

Theorems & Definitions (4)

• Remark 1
• Remark 2
• Remark 3
• Remark 4