Intelligent Switching for Reset-Free RL

In the real world, the strong episode resetting mechanisms that are needed to trainagents in simulation are unavailable. The resetting assumption limits the potentialof reinforcement learning in the real world, as providing resets to an agent usuallyrequires the creation of additional handcrafted mechanisms or human interventions.Recent work aims to train agents (forward) with learned resets by constructinga second (backward) agent that returns the forward agent to the initial state. Wefind that the termination and timing of the transitions between these two agentsare crucial for algorithm success. With this in mind, we create a new algorithm,Reset Free RL with Intelligently Switching Controller (RISC) which intelligentlyswitches between the two agents based on the agent’s confidence in achieving itscurrent goal. Our new method achieves state-of-the-art performance on severalchallenging environments for reset-free RL.