On Time-Indexing as Inductive Bias in Deep RL for Sequential Manipulation Tasks

The article presents a study on how to improve skill acquisition in sequential manipulation tasks using Deep Reinforcement Learning (Deep RL). The authors propose a policy architecture where different action heads are executed sequentially for fixed durations. This allows the learning of primitive skills such as reaching and grasping. The method was tested on Metaworld tasks, where it outperformed standard policy learning methods, demonstrating its potential for better skill acquisition. The study argues that time-indexing can serve as a useful indicator for skill selection in sequential tasks.

Publication date: 5 Jan 2024
Project Page: unknown
Paper: https://arxiv.org/pdf/2401.01993