Hierarchical Entity-centric Reinforcement Learning with Factored Subgoal Diffusion

TL;DR

HECRL tackles long-horizon, multi-entity goal-conditioned RL in offline settings by combining a value-based low-level agent with a high-level, entity-factored Subgoal Diffuser. The diffusion-based subgoal generator is trained on offline data and, at test time, selects subgoals that are within a computed competence radius using the value function, enabling efficient, modular planning without altering the underlying RL algorithm. Factored subgoals produced by the diffusion model bias the agent toward sparse changes to state factors, improving tractability in combinatorial, image-based observation spaces. Empirically, HECRL yields substantial gains over baselines on several long-horizon multi-object tasks, including more than a 150% increase in success rate on the hardest image-based PPP-Cube task, and shows non-trivial zero-shot generalization as the number of entities grows, with limitations tied to the quality of the factored representations and the need for robust auto-encoding of scene factors.

Abstract

We propose a hierarchical entity-centric framework for offline Goal-Conditioned Reinforcement Learning (GCRL) that combines subgoal decomposition with factored structure to solve long-horizon tasks in domains with multiple entities. Achieving long-horizon goals in complex environments remains a core challenge in Reinforcement Learning (RL). Domains with multiple entities are particularly difficult due to their combinatorial complexity. GCRL facilitates generalization across goals and the use of subgoal structure, but struggles with high-dimensional observations and combinatorial state-spaces, especially under sparse reward. We employ a two-level hierarchy composed of a value-based GCRL agent and a factored subgoal-generating conditional diffusion model. The RL agent and subgoal generator are trained independently and composed post hoc through selective subgoal generation based on the value function, making the approach modular and compatible with existing GCRL algorithms. We introduce new variations to benchmark tasks that highlight the challenges of multi-entity domains, and show that our method consistently boosts performance of the underlying RL agent on image-based long-horizon tasks with sparse rewards, achieving over 150% higher success rates on the hardest task in our suite and generalizing to increasing horizons and numbers of entities. Rollout videos are provided at: https://sites.google.com/view/hecrl

Figures (14)

• Figure 1: Sceneimage-based factored subgoals. Bottom row: environment image observations at the time of subgoal generation. Top row: Reconstructions of the generated latent factored subgoals. Rightmost image: goal image observation. Circles and arrows: circles on the top row highlight factors (excluding the arm) that are modified by the subgoals compared to the current observation in the image below them. Arrows leading to circles on the bottom row highlight the factors that are manipulated by the low-level RL policy, showing that the subgoals were achieved. Subgoal reconstructions only include the foreground of the scene captured by the Deep Latent Particles (DLP, daniel2022unsupervised) representation. See App. for details on the task \ref{['apdx:envs']} and the DLP model \ref{['apdx:related:dlp']}.
• Figure 2: Left: subgoal sampling illustration (Alg. \ref{['alg:subgoal']} lines $2$--$4$). Robot - agent, Dashed circle - competence radius, Flag - goal, Red triangle - discarded subgoal, Green circle - filtered subgoal, Yellow star - chosen subgoal, Background color gradient - value landscape. Right: factored subgoal illustration. Factored subgoals make it easy to modify small subsets of entities which simplifies the subtask when factors are independently controllable. The subgoal images depict two states that are roughly the same distance from the initial state, where the top requires moving only the blue cube while the bottom requires moving all three.
• Figure 3: Environment suite. PPP-Cube: Pick, Place and Push cubes to goal positions. Stack-Cube: Pick-and-place cube stacking. Scene: Manipulate objects (drawer, window, button and cube) to goal configurations. Push-Tetris: Push blocks to goal positions and orientations.
• Figure 4: PPP-Cube and Push-Tetrisimage-based factored subgoals. Bottom row: environment image observations at the time of subgoal generation. Positions of latent particles are plotted on the image. Top row: DLP reconstructions of the generated latent subgoals. Rightmost image: goal image observation. Circles and arrows: circles on the top row highlight factors (excluding the agent) that are modified by the subgoals compared to the current observation in the image below them. Arrows leading to circles on the bottom row highlight the factors that are manipulated by the low-level RL policy, showing that the subgoals were achieved.
• Figure 5: Hyperparameter sensitivity study. Success rates of our method on PPP-Cube(Image). A single hyperparameter is varied in each plot while the others are unchanged compared to the ones used for the main results, which we refer to as default. Purple stars - default values, error bars - standard deviations, dashed lines and shaded regions in green - mean and standard deviation of the EC-IQL baseline without subgoals. Left: varying $K$, the Subgoal Diffuser training distribution timestep difference between state and subgoal. Center: varying $T_{sg}$, the test-time subgoal rollout timesteps. Right: varying $N$, the number of test-time diffusion subgoals sampled before filtering.