Autonomous Reinforcement Learning: Formalism and Benchmarking

Abstract

Reinforcement learning (RL) provides a naturalistic framing for learning through trial and error, which is appealing both because of its simplicity and effectiveness and because of its resemblance to how humans and animals acquire skills through experience. However, real-world embodied learning, such as that performed by humans and animals, is situated in a continual, non-episodic world, whereas common benchmark tasks in RL are episodic, with the environment resetting between trials to provide the agent with multiple attempts. This discrepancy presents a major challenge when attempting to take RL algorithms developed for episodic simulated environments and run them on real-world platforms, such as robots. In this paper, we aim to address this discrepancy by laying out a framework for Autonomous Reinforcement Learning (ARL): reinforcement learning where the agent not only learns through its own experience, but also contends with lack of human supervision to reset between trials. We introduce a simulated benchmark EARL 1 around this framework, containing a set of diverse and challenging simulated tasks reflective of the hurdles introduced to learning when only a minimal reliance on extrinsic intervention can be assumed. We show that standard approaches to episodic RL and existing approaches struggle as interventions are minimized, underscoring the need for developing new algorithms for reinforcement learning with a greater focus on autonomy.

Don't Start From Scratch: Leveraging Prior Data to Automate Robotic Reinforcement Learning

“…Our system aims to utilize prior data to autonomously learn robotic skills, leveraging prior experience to both accelerate the learning of a new behavior and the process of learning how to reset between attempts. Prior work has tackled this "reset-free" learning problem [30] in a number of ways. Some prior work uses a curriculum-based approach, which relies on the observation that when learning several tasks simultaneously, some tasks reset others, thus forming a curriculum [31,32,33,4,34].…”

“…1.) In both settings, when the robot is practicing on the target task τ * , it is provided with minimal external resets, in accordance with the autonomous RL paradigm [30]. The learned agent is then evaluated on the same task τ * , but this time the environment is reset after every episode.…”