Imagine Beyond! Distributionally Robust Auto-Encoding for State Space Coverage in Online Reinforcement Learning

Goal-Conditioned Reinforcement Learning (GCRL) enables agents to autonomously acquire diverse behaviors, but faces major challenges in visual environments due to high-dimensional, semantically sparse observations. In the online setting, where agents learn representations while exploring, the latent space evolves with the agent's policy, to capture newly discovered areas of the environment. However, without incentivization to maximize state coverage in the representation, classical approaches based on auto-encoders may converge to latent spaces that over-represent a restricted set of states frequently visited by the agent. This is exacerbated in an intrinsic motivation setting, where the agent uses the distribution encoded in the latent space to sample the goals it learns to master. To address this issue, we propose to progressively enforce distributional shifts towards a uniform distribution over the full state space, to ensure a full coverage of skills that can be learned in the environment. We introduce DRAG (Distributionally Robust Auto-Encoding for GCRL), a method that combines the $\beta$-VAE framework with Distributionally Robust Optimization. DRAG leverages an adversarial neural weighter of training states of the VAE, to account for the mismatch between the current data distribution and unseen parts of the environment. This allows the agent to construct semantically meaningful latent spaces beyond its immediate experience. Our approach improves state space coverage and downstream control performance on hard exploration environments such as mazes and robotic control involving walls to bypass, without pre-training nor prior environment knowledge.

Stein Variational Goal Generation For Reinforcement Learning in Hard Exploration Problems

Multi-goal Reinforcement Learning has recently attracted a large amount of research interest. By allowing experience to be shared between related training tasks, this setting favors generalization for new tasks at test time, whenever some smoothness exists in the considered representation space of goals. However, in settings with discontinuities in state or goal spaces (e.g. walls in a maze), a majority of goals are difficult to reach, due to the sparsity of rewards in the absence of expert knowledge. This implies hard exploration, for which some curriculum of goals must be discovered, to help agents learn by adapting training tasks to their current capabilities. Building on recent automatic curriculum learning techniques for goal-conditioned policies, we propose a novel approach: Stein Variational Goal Generation (SVGG), which seeks at preferably sampling new goals in the zone of proximal development of the agent, by leveraging a learned model of its abilities, and a goal distribution modeled as particles in the exploration space. Our approach relies on Stein Variational Gradient Descent to dynamically attract the goal sampling distribution in areas of appropriate difficulty. We demonstrate the performances of the approach, in terms of success coverage in the goal space, compared to recent state-of-the-art RL methods for hard exploration problems.