DMC-VB: A Benchmark for Representation Learning for Control with Visual Distractors

TL;DR

The paper introduces DMC-VB, a comprehensive benchmark dataset for evaluating the robustness of offline reinforcement learning agents in continuous control tasks with visual distractors, highlighting challenges in representation learning from visual inputs.

Contribution

The paper presents a new large-scale dataset and benchmark for assessing representation learning methods in control tasks with visual distractors, including diverse tasks and systematic evaluation protocols.

Findings

Pretrained representations do not improve policy learning on DMC-VB.

A significant gap exists between policies learned on pixel observations and on true states.

Policy learning benefits from pretrained representations when expert data is limited, especially with suboptimal or stochastic hidden goal data.

Abstract

Learning from previously collected data via behavioral cloning or offline reinforcement learning (RL) is a powerful recipe for scaling generalist agents by avoiding the need for expensive online learning. Despite strong generalization in some respects, agents are often remarkably brittle to minor visual variations in control-irrelevant factors such as the background or camera viewpoint. In this paper, we present theDeepMind Control Visual Benchmark (DMC-VB), a dataset collected in the DeepMind Control Suite to evaluate the robustness of offline RL agents for solving continuous control tasks from visual input in the presence of visual distractors. In contrast to prior works, our dataset (a) combines locomotion and navigation tasks of varying difficulties, (b) includes static and dynamic visual variations, (c) considers data generated by policies with different skill levels, (d) systematically returns pairs of state and pixel observation, (e) is an order of magnitude larger, and (f) includes tasks with hidden goals. Accompanying our dataset, we propose three benchmarks to evaluate representation learning methods for pretraining, and carry out experiments on several recently proposed methods. First, we find that pretrained representations do not help policy learning on DMC-VB, and we highlight a large representation gap between policies learned on pixel observations and on states. Second, we demonstrate when expert data is limited, policy learning can benefit from representations pretrained on (a) suboptimal data, and (b) tasks with stochastic hidden goals. Our dataset and benchmark code to train and evaluate agents are available at: https://github.com/google-deepmind/dmc_vision_benchmark.