Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning Latent Dynamics for Planning from Pixels
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 |
|---|---|---|---|
| Use PlaNet instead of D4PG | Cheetah Run · Cheetah Run | final performance: 26% | Helped |
| Use model-based PlaNet rather than D4PG | continuous control tasks from pixels · DeepMind control suite tasks | environment episodes to comparable performance: 200× fewer episodes | Helped |
| Add both deterministic and stochastic transition paths (ablation) | PlaNet dynamics model · six control tasks | test performance | Helped |
| Remove stochastic transitions (use a deterministic GRU) (ablation) | PlaNet · six control tasks | test performance | Hurt |
| Remove deterministic recurrent transitions (use a stochastic SSM) (ablation) | PlaNet · six control tasks | test performance | Hurt |
| Collect experience through planning instead of random actions (ablation) | PlaNet · six control tasks | test performance | Helped |
| Replace iterative CEM refinement with selecting the best of 1000 sampled sequences (ablation) | PlaNet · six control tasks | test performance | Hurt |
| Add latent overshooting (ablation) | DRNN and other tested dynamics models · control tasks | test performance | Helped |
| Add latent overshooting (ablation) | RSSM · control tasks | test performance | Hurt |
| Replace hard ReLU with smooth ELU activations (ablation) | purely stochastic model · control tasks | test performance | Helped |
| Replace hard ReLU with smooth ELU activations (ablation) | RSSM · control tasks | test performance | No clear effect |
| Train one agent on all six tasks rather than separate agents (ablation) | PlaNet · six control tasks | learning speed | Hurt |
Closest papers by meaning
- HaM-World: Soft-Hamiltonian World Models with Selective Memory for Planning (NeurIPS 2026 · similarity 0.88)
- Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models (NeurIPS 2025 · similarity 0.88)
- Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.88)
- Predictive but Not Plannable: RC-aux for Latent World Models (NeurIPS 2026 · similarity 0.87)
- Latent Plan Transformer for Trajectory Abstraction: Planning as Latent Space Inference (NeurIPS 2024 · similarity 0.87)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 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.