Data-Efficient Multirobot, Multitask Transfer Learning for Trajectory Tracking

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 paper, we introduce a multi-robot, multi-task 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, multi-robot transfer learning framework is based on a combined L1 adaptive control and 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 multi-task 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.

Data-Driven Hierarchical Predictive Learning in Unknown Environments

“…Critically, the target set must belong to the safe set (20). This ensures that once the system has reached the target set, there will always exist at least one feasible input (the safety control (21)) that allows the system to satisfy all state constraints. Given a strategy set (19), we find a corresponding lifted target set:…”

Knowledge Transfer Between Robots with Similar Dynamics for High-Accuracy Impromptu Trajectory Tracking

“…This offline mapping can then be used to translate the policies trained on the source robot to the policies for the target robot [2], or map the data collected on the source robot to the target robot for model learning [3]. Extension hereto [4], [16] derive an optimal mapping for data transfer across robots from a control theory perspective. Other related work aims to learn and exploit a common feature space between the source and target robots while performing similar tasks [5], [6].…”

“…The collected data can often be compactly represented by parametric regression techniques [8], [9]. For the source system (16), we adopt the DNN module from [9] and transfer this inverse module to enhance the target system (17) with the proposed online learning approach. The DNN module of the source system is a 3-layer feedforward network with 20 hyperbolic tangent neurons in each hidden layer.…”

A New Cloud Robots Training Method Using Cooperative Learning

“…To this day, many researches have been done on the learning and sharing of knowledge among robots. For example, a multi-robot, multi-tasking learning framework, after a robot has passed the demonstration learning task, lessons learned can be moved to other robots and used to perform another task [3]. A learning framework of adaptive manipulative skills from human to robot to facilitate robot skill generalization is described in [4].…”