Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Dream to Control: Learning Behaviors by Latent Imagination
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Learn actions and state values in latent imagination, propagating analytic value gradients through the learned dynamics | Dreamer · DeepMind Control Suite (20-task average) | score: 332.97 -> 823.39 | Helped |
| Use Dreamer instead of D4PG | DeepMind Control Suite (20-task average) · DeepMind Control Suite | score: 786.32 -> 823.39 | Helped |
| Use Dreamer rather than online planning with PlaNet | DeepMind Control Suite · DeepMind Control Suite | training time per 10^6 environment steps: 11 hours -> 3 hours | Helped |
| Add state-value prediction to action learning in imagination (ablation) | Dreamer · DeepMind Control Suite, different imagination horizons | task performance and robustness to horizon length | Helped |
| Replace learned actions in imagination with online planning (ablation) | Dreamer world model · DeepMind Control Suite, long-horizon tasks | task performance | Hurt |
| Remove the value model and optimize only rewards within the imagination horizon (ablation) | Dreamer · DeepMind Control Suite, long-horizon tasks | task performance | Hurt |
| Use pixel reconstruction instead of contrastive estimation for representation learning (ablation) | Dreamer · DeepMind Control Suite | task performance | Helped |
| Use contrastive estimation instead of pure reward prediction for representation learning (ablation) | Dreamer · DeepMind Control Suite | task performance | Helped |
| Use pure reward prediction rather than image reconstruction or contrastive estimation (ablation) | Dreamer · DeepMind Control Suite | task performance | Hurt |
| Vary action repeat; the paper reports that R=2 works best across tasks (ablation) | Dreamer · DeepMind Control Suite | task performance | Mixed |
| Add latent overshooting during world-model learning (ablation) | Dreamer · DeepMind Control Suite | task performance | No clear effect |
| Add an entropy bonus to the action model (ablation) | Dreamer · DeepMind Control Suite | task performance | No clear effect |
Closest papers by meaning
- Open-World Reinforcement Learning over Long Short-Term Imagination (ICLR 2025 · similarity 0.90)
- Dynamics-Aligned Latent Imagination in Contextual World Models for Zero-Shot Generalization (NeurIPS 2025 · similarity 0.87)
- Zero-shot World Models via Search in Memory (NeurIPS 2025 · similarity 0.87)
- Dr. Strategy: Model-Based Generalist Agents with Strategic Dreaming (ICML 2024 · similarity 0.87)
- RIG: Synergizing Reasoning and Imagination in End-to-End Generalist Policy (ICLR 2026 · similarity 0.86)
- Learning Latent Dynamic Robust Representations for World Models (ICML 2024 · similarity 0.86)
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.