Original title: Enhancing Robotic Manipulation: Harnessing the Power of Multi-Task Reinforcement Learning and Single Life Reinforcement Learning in Meta-World
Authors: Ghadi Nehme, Ishan Sabane, Tejas Y. Deo
The article dives into the challenge of training robots to handle multiple tasks efficiently. Existing multi-task reinforcement learning (RL) algorithms excel in similar environments but struggle when faced with diverse scenarios. Q-Weighted Adversarial Learning (QWALE) addresses this by training for specific tasks, hindering generalization. This study focuses on empowering a robotic arm to master seven tasks in the Meta World environment. It uses a multi-task soft actor-critic (MT-SAC) to train the arm and then utilizes this trained model as prior data for a single-life RL algorithm, MT-QWALE. Assessing its performance across novel positions, MT-QWALE outshines MT-SAC. Even after concealing the final goal position, MT-QWALE completes tasks in slightly more steps. This comparison highlights MT-QWALE’s efficacy in handling varied scenarios, providing insights into its potential for multi-task robotic learning.
Original article: https://arxiv.org/abs/2311.12854