DEAR: Disentangled Environment and Agent Representations for Reinforcement Learning without Reconstruction

TL;DR

A novel method, Disentangled Environment and Agent Representations (DEAR), that uses the segmentation mask of the agent as supervision to learn disentangled representations of the environment and the agent through feature separation constraints, which surpasses state-of-the-art methods in sample efficiency.

Abstract

Reinforcement Learning (RL) algorithms can learn robotic control tasks from visual observations, but they often require a large amount of data, especially when the visual scene is complex and unstructured. In this paper, we explore how the agent's knowledge of its shape can improve the sample efficiency of visual RL methods. We propose a novel method, Disentangled Environment and Agent Representations (DEAR), that uses the segmentation mask of the agent as supervision to learn disentangled representations of the environment and the agent through feature separation constraints. Unlike previous approaches, DEAR does not require reconstruction of visual observations. These representations are then used as an auxiliary loss to the RL objective, encouraging the agent to focus on the relevant features of the environment. We evaluate DEAR on two challenging benchmarks: Distracting DeepMind control suite and Franka Kitchen manipulation tasks. Our findings demonstrate that DEAR surpasses state-of-the-art methods in sample efficiency, achieving comparable or superior performance with reduced parameters. Our results indicate that integrating agent knowledge into visual RL methods has the potential to enhance their learning efficiency and robustness.

Figures (6)

• Figure 1: Robust representations of the visual scene should disentangle agent representations from the irrelevant environment information (e.g., changing background). Agent mask can extract agent representations from state observations, while representations not included in agent features can be pushed far apart.
• Figure 2: DEAR architecture: An agent mask is used as a supervision to encode agent features $z_A$. Additionally, the latent distance between $z_A$ and $z_E$ is maximized to encode environment features in $z_E$. A concatenation of $z_A$ and $z_E$ is used as the input to the RL policy $\pi$.
• Figure 3: Graphical model for (left) SEAR, and DEAR (right)
• Figure 4: Different environments used for validation. (top row) RGB images that are input to the agent (bottom row) agent segmentation masks.
• Figure 6: Training curves of different methods on three distracting DM control environments and two Franka kitchen tasks. DEAR outperforms or matches other baselines on these tasks. Each curve is an average of 5 different seeds.

Theorems & Definitions (1)

• Conjecture 1