Research map / Reinforcement learning / Visual and representation RL
Task-Agnostic Dynamics Priors for Deep Reinforcement Learning
Yilun Du, Karthic Narasimhan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SpatialNet rather than RCNet or ConvLSTM for noisy-input prediction | Physics prediction · Physics prediction, ε=0.5 | MSE: 0.0268 (RCNet), 0.0072 (ConvLSTM) -> 0.0062 | Improved |
| Use SpatialNet for generalization to smaller objects moving faster | Video prediction · Unseen smaller and faster objects | MSE: 0.0115 (RCNet), 0.0067 (ConvLSTM) -> 0.0039 | Improved |
| Use SpatialNet for prediction at a larger input resolution | Video prediction · 168x168 input images | MSE: 0.0060 -> 0.0042 | Improved |
| Initialize SpatialNet with physics-dataset pretraining | Atari video prediction · Assault | MSE: 0.00522 -> 0.00477 | Improved |
| Initialize SpatialNet with physics-dataset pretraining | Atari video prediction · DemonAttack | MSE: 0.00562 -> 0.00433 | Improved |
| Initialize SpatialNet with physics-dataset pretraining | Atari video prediction · Enduro | MSE: 0.00411 -> 0.00576 | Hurt |
| Initialize SpatialNet with physics-dataset pretraining | Atari video prediction · Pong | MSE: 0.00584 -> 0.00636 | Hurt |
| Initialize SpatialNet with physics-dataset pretraining | Atari video prediction · 10 Atari environments | MSE: 12.9% decrease in MSE error | Mixed |
| Use D2A instead of PPO | Atari policy learning · Stochastic Atari environments | Atari score | Mixed |
| Use D2A instead of PPO | Atari policy learning · Asteroids | Atari score: 1321.0 ± 233.5 -> 2098.4 ± 102.0 | Improved |
| Use D2A instead of PPO | Atari policy learning · Boxing | Atari score: 54.3 ± 2.5 -> 8.9 ± 20.0 | Hurt |
Closest papers by meaning
- Context-aware Dynamics Model for Generalization in Model-Based Reinforcement Learning (ICML 2020 · similarity 0.89)
- The Surprising Ineffectiveness of Pre-Trained Visual Representations for Model-Based Reinforcement Learning (NeurIPS 2024 · similarity 0.88)
- Behavior-agnostic Task Inference for Robust Offline In-context Reinforcement Learning (ICML 2025 · similarity 0.87)
- DrM: Mastering Visual Reinforcement Learning through Dormant Ratio Minimization (ICLR 2024 · similarity 0.87)
- Towards General-Purpose Model-Free Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
