Data-Efficient Multirobot, Multitask Transfer Learning for Trajectory Tracking

TL;DR

This paper presents a data-efficient transfer learning framework for multirobot trajectory tracking, enabling robots to learn new tasks from few demonstrations by combining adaptive control and iterative learning control, validated on quadrotors.

Contribution

It introduces a novel multirobot, multitask transfer learning method that leverages control theory to enable accurate trajectory tracking with minimal data across different robots.

Findings

Reduces first-iteration tracking error by 74% using transfer learning.

Effective across different quadrotor platforms and trajectories.

Combines adaptive control with iterative learning for robust transfer.

Abstract

Transfer learning has the potential to reduce the burden of data collection and to decrease the unavoidable risks of the training phase. In this letter, we introduce a multirobot, multitask transfer learning framework that allows a system to complete a task by learning from a few demonstrations of another task executed on another system. We focus on the trajectory tracking problem where each trajectory represents a different task, since many robotic tasks can be described as a trajectory tracking problem. The proposed multirobot transfer learning framework is based on a combined $\mathcal{L}_1$ adaptive control and an iterative learning control approach. The key idea is that the adaptive controller forces dynamically different systems to behave as a specified reference model. The proposed multitask transfer learning framework uses theoretical control results (e.g., the concept of vector relative degree) to learn a map from desired trajectories to the inputs that make the system track these trajectories with high accuracy. This map is used to calculate the inputs for a new, unseen trajectory. Experimental results using two different quadrotor platforms and six different trajectories show that, on average, the proposed framework reduces the first-iteration tracking error by 74% when information from tracking a different single trajectory on a different quadrotor is utilized.