Research map / Reinforcement learning / Visual and representation RL
Efficient Planning in a Compact Latent Action Space
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 |
|---|---|---|---|
| add latent-action planning (state-conditional VQ-VAE with L=3 latent codes) instead of raw-action Trajectory Transformer planning | TAP · Adroit (24-DoF hand tasks) | normalised score mean without expert: 19.6 vs 6.1 for TT | Helped |
| latent-action planning (TAP) vs. raw-action planning (TT) | locomotion (D4RL gym tasks) · D4RL locomotion mean | normalised score: 82.5 vs 82.2 | No clear effect |
| use TAP instead of model-free actor-critic baselines | Adroit · Adroit all settings mean | normalised score: 51.9 vs 40.3 for CQL | Helped |
| reduce latent steps per code from L=3 to L=1 (ablation) | TAP · D4RL locomotion mean | normalised score: 60.6 vs 82.5 | Hurt |
| use L=3 latent steps per code (vs L=2) (ablation) | TAP · D4RL locomotion mean | normalised score: 82.5 vs 73.6 | Helped |
| increase latent steps per code from L=3 to L=4 (ablation) | TAP · D4RL locomotion mean | normalised score: 82.1 vs 82.5 | No clear effect |
| reduce planning horizon to a single latent variable (3 steps) (ablation) | TAP · D4RL locomotion mean | normalised score: 71.9 vs 82.5 at horizon 15 | Hurt |
| OOD penalty threshold beta=1e-5 (penalty effectively inactive) (ablation) | TAP · D4RL locomotion mean | normalised score: 57.6 vs 82.5 | Hurt |
| OOD penalty threshold beta=0.5 (penalty always active) (ablation) | TAP · D4RL locomotion mean | normalised score: 65.8 vs 82.5 | Hurt |
| beam search instead of direct sampling from prior (2048 samples) (ablation) | TAP · D4RL locomotion mean | normalised score: 82.5 vs 78.3 | Helped |
| uniform latent code sampling instead of learned prior sampling (ablation) | TAP · D4RL locomotion mean | normalised score: 51.3 vs 78.3 (prior sampling) | Hurt |
| remove state conditioning from decoder (unconditional decoder) (ablation) | TAP · D4RL locomotion mean | normalised score: 21.4 to 22.4 vs 82.5 for conditional decoder | Hurt |
Closest papers by meaning
- Scalable Decision-Making in Stochastic Environments through Learned Temporal Abstraction (ICLR 2025 · similarity 0.90)
- Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.87)
- LaST-R1: Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning (NeurIPS 2026 · similarity 0.87)
- Efficient Planning with Latent Diffusion (ICLR 2024 · similarity 0.86)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.86)
- PcLast: Discovering Plannable Continuous Latent States (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.