Research map / Image generation / Fast diffusion sampling

Tree-Guided Diffusion Planner

NeurIPS 2025 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TDP, a training-free zero-shot test-time planner that wraps a pretrained diffusion planner in a bi-level tree search (particle-guided parent trajectories plus gradient-guided fast-denoised child refinements) to handle non-convex, non-differentiable, and multi-reward guidance.

Problem

Test-time guided planning with pretrained diffusion planners fails when guide functions are non-convex, non-differentiable (e.g., intermediate-goal constraints), or multi-reward, because standard gradient guidance gets trapped in local optima and in-distribution trajectories.

Approach

Frames test-time planning as tree search with a bi-level sampling process: fixed-potential particle guidance on control states produces diverse parent trajectories, and fast partial re-denoising from random branch points with gradient guidance on observation states produces child trajectories; the best leaf is selected by the true guide score.

Key results

• On Maze2D gold-picking, TDP reaches 43.7 single-task and 72.4 multi-task average versus 21.1 and 44.9 for MCSS.
• On KUKA PnWP, TDP scores 66.81 versus 39.72 for AdaptDiffuser and 35.69 for MCSS.
• On AntMaze multi-goal, TDP finds 66.1 goals versus 61.3 for MCSS and reduces timesteps per goal to 558.4 from 612.1.

Measured effects reported in the paper

Surprising findings

• Learned particle guidance gives only marginal gains over fixed particle guidance (e.g., PnWP 67.23 vs 66.81), so the fixed potential is nearly as good.
• Conditional particle guidance in parent branching can bias samples toward local optima, which hurts performance without sub-tree expansion.

Limitations

• Bi-level sampling adds computational cost from pairwise trajectory distance computation and extra fast denoising.
• Gradient guidance strength alpha_g must still be tuned per task.
• Maze2D gold-picking uses an approximate distance guide for gradient guidance, with the true guide used only for selection.

Experimental setup

Datasets: Maze2D PDDLStream block stacking demonstrations AntMaze

Benchmarks: Maze2D gold-picking KUKA PnP (stack) KUKA PnP (place) KUKA PnWP AntMaze multi-goal exploration

Metrics: mean performance score number of found goals sequence match score timesteps per goal

Baselines: Diffuser Diffuser-gamma (TAT) AdaptDiffuser MCSS MCSS+SS TDP w/o child TDP w/o PG

Builds on: Diffuser particle guidance Trajectory Aggregation Tree Monte Carlo tree diffusion classifier guidance

Closest papers by meaning

• Task-Agnostic Pre-training and Task-Guided Fine-tuning for Versatile Diffusion Planner (ICML 2025 · similarity 0.89)
• Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.89)
• Resisting Stochastic Risks in Diffusion Planners with the Trajectory Aggregation Tree (ICML 2024 · similarity 0.88)
• Prior-Guided Diffusion Planning for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
• Compositional Monte Carlo Tree Diffusion for Extendable Planning (NeurIPS 2025 · similarity 0.87)
• Fast Monte Carlo Tree Diffusion: 100× Speedup via Parallel and Sparse Planning (NeurIPS 2025 · similarity 0.87)