Episodic Curiosity through Reachability

TL;DR

This paper introduces a novel episodic curiosity method using reachability-based memory comparisons to enhance exploration in sparse reward environments, outperforming existing methods in visual 3D navigation and locomotion tasks.

Contribution

The paper presents a new curiosity approach leveraging episodic memory and reachability to better measure novelty, overcoming prior issues like the 'couch-potato' problem.

Findings

Outperforms ICM in ViZDoom and DMLab navigation tasks.

Enables MuJoCo ant to learn locomotion from first-person curiosity.

Effective in visually rich 3D environments.

Abstract

Rewards are sparse in the real world and most of today's reinforcement learning algorithms struggle with such sparsity. One solution to this problem is to allow the agent to create rewards for itself - thus making rewards dense and more suitable for learning. In particular, inspired by curious behaviour in animals, observing something novel could be rewarded with a bonus. Such bonus is summed up with the real task reward - making it possible for RL algorithms to learn from the combined reward. We propose a new curiosity method which uses episodic memory to form the novelty bonus. To determine the bonus, the current observation is compared with the observations in memory. Crucially, the comparison is done based on how many environment steps it takes to reach the current observation from those in memory - which incorporates rich information about environment dynamics. This allows us to overcome the known "couch-potato" issues of prior work - when the agent finds a way to instantly gratify itself by exploiting actions which lead to hardly predictable consequences. We test our approach in visually rich 3D environments in ViZDoom, DMLab and MuJoCo. In navigational tasks from ViZDoom and DMLab, our agent outperforms the state-of-the-art curiosity method ICM. In MuJoCo, an ant equipped with our curiosity module learns locomotion out of the first-person-view curiosity only.