Learning Latent Dynamic Robust Representations for World Models
TLDR
Improves visual MBRL world models by learning robust representations via spatio-temporal masking, bisimulation, and HRSSM, achieving better performance on visually complex control tasks.
Reasoning
The paper presents a novel combination of spatio-temporal masking, bisimulation, and a hybrid recurrent state-space model to address exogenous noise in visual MBRL. Strengths include clear problem motivation and empirical gains on challenging tasks with real-world distractors. Weaknesses are the narrow focus on specific control tasks and lack of broader theoretical analysis or comparison to other robust representation methods.
Read-first score
Read-first score 59.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 34.
Field roles
Rank sensitivity
Stability: volatile; rank range: 256.
Keyword Scores
Deep Analysis
Innovations
- Spatio-temporal masking strategy combined with bisimulation principle and latent reconstruction to capture endogenous task-specific features and filter exogenous noise
- Hybrid Recurrent State-Space Model (HRSSM) to address instabilities in joint training of representations, dynamics, and policy
Methodology
The paper proposes a visual model-based reinforcement learning approach that applies a spatio-temporal masking strategy and bisimulation principle along with latent reconstruction to learn robust representations that ignore exogenous noise. It introduces a Hybrid Recurrent State-Space Model (HRSSM) structure to stabilize joint training of representations, dynamics, and policy. The method is evaluated on visually complex control tasks from Maniskill with exogenous distractors from the Matterport environment.
Key Results
The empirical evaluation demonstrates significant performance improvements over existing methods, including Dreamer, in visually complex control tasks with exogenous distractors.