R2-Dreamer: Redundancy-Reduced World Models without Decoders or Augmentation
TLDR
R2-Dreamer introduces a redundancy-reduction objective for decoder-free MBRL, achieving faster training and competitive performance without data augmentation.
Reasoning
Strengths include a novel self-supervised internal regularizer that eliminates the need for data augmentation, leading to faster training. Weaknesses are limited evaluation to simulated environments (DMC, Meta-World) and lack of real-world experiments.
Read-first score
Read-first score 73.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.
Field roles
Rank sensitivity
Stability: volatile; rank range: 114.
Keyword Scores
Deep Analysis
Innovations
- Decoder-free MBRL with self-supervised redundancy-reduction objective as internal regularizer, eliminating need for data augmentation
- Integration of Barlow Twins-inspired objective into world models for representation learning
- Faster training (1.59x) compared to DreamerV3 while maintaining competitive performance
Methodology
R2-Dreamer is a decoder-free model-based reinforcement learning framework that uses a redundancy-reduction objective inspired by Barlow Twins as an internal regularizer to prevent representation collapse without data augmentation. The model is trained on image-based tasks from DeepMind Control Suite and Meta-World, and evaluated against DreamerV3 and TD-MPC2 baselines.
Key Results
R2-Dreamer achieves competitive performance with DreamerV3 and TD-MPC2 on standard benchmarks, trains 1.59x faster than DreamerV3, and yields substantial gains on DMC-Subtle with tiny task-relevant objects.