Research map / Reinforcement learning / Visual and representation RL
Diffused Task-Agnostic Milestone Planner
Mineui Hong, Minjae Kang, Songhwai Oh
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 |
|---|---|---|---|
| Plan temporally sparse latent milestones and follow them with a goal-conditioned actor | DTAMP versus offline RL and sequence-modeling baselines · D4RL Antmaze | score: 85.0$\pm$4.84 vs 61.7 | Improved |
| Plan latent milestones rather than whole trajectories | DTAMP versus sequence modeling baselines · Antmaze long-horizon tasks | score: DTAMP 93.3$\pm$0.94; DT, TT, DD 0.0 | Improved |
| Use diffusion-generated milestones for varying goals | DTAMP versus multi-goal baselines · Multi-goal Antmaze average | score: 70.3$\pm$5.86 vs 49.2 | Improved |
| Re-plan the remaining milestones after reaching a milestone | DTAMP · CALVIN average | success rate: 54.8$\pm$4.37 -> 62.7$\pm$3.60 | Improved |
| Add reconstruction loss when training the encoder (ablation) | DTAMP · CALVIN average | success rate: 50.9 -> 54.8 | Improved |
| Add goal-conditioned imitation learning loss alongside reconstruction loss (ablation) | DTAMP · CALVIN average | success rate: 49.1 -> 54.8 | Improved |
| Use skill-based policy (ablation) | DTAMP · CALVIN average | success rate: 31.8 -> 54.8 | Improved |
| Add reconstruction loss when training the encoder (ablation) | DTAMP · CALVIN 1 task | success rate: 86.7 -> 82.4 | Hurt |
| Guide diffusion toward smaller temporal intervals between milestones (ablation) | DTAMP milestone planner · Simplified open Antmaze | timesteps to reach goal | Improved |
| Apply minimum-temporal-distance diffusion guidance (ablation) | DTAMP milestone planner · Simplified open Antmaze | success rate | No clear effect |
| Train an encoder using both reconstruction and goal-conditioned imitation learning losses (ablation) | DTAMP · CALVIN multi-task planning | success rate: 86.7 vs 82.4; 43.3 vs 52.4; 17.2 vs 29.7 | Mixed |
| Use a feedback actor and plan milestones once at episode start | DTAMP versus Diffuser · maze2d-umaze-v1 under increasing stochasticity | normalized score: DTAMP 97.6, 99.0, 99.4; Diffuser 61.9, 29.6, 18.5 | Improved |
Closest papers by meaning
- State-Covering Trajectory Stitching for Diffusion Planners (NeurIPS 2025 · similarity 0.90)
- Simple Hierarchical Planning with Diffusion (ICLR 2024 · similarity 0.90)
- Task-Agnostic Pre-training and Task-Guided Fine-tuning for Versatile Diffusion Planner (ICML 2025 · similarity 0.88)
- Efficient Planning with Latent Diffusion (ICLR 2024 · similarity 0.88)
- GenPlan: Generative Sequence Models as Adaptive Planners (AAAI 2025 · similarity 0.88)
- Generative Trajectory Stitching through Diffusion Composition (NeurIPS 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.
