WorldPlanner: Monte Carlo Tree Search and MPC with Action-Conditioned Visual World Models

arXiv 25.11 2025 65 method, application

TLDR

A model-based approach using action-conditioned visual world models with MCTS and MPC for robotic planning, validated on real-world tasks.

Reasoning

The paper presents a novel integration of world models with Monte Carlo Tree Search and Model Predictive Control, demonstrating improvements over behavior cloning on real robotic tasks. Strengths include real-world validation and a diffusion-based action sampler to reduce hallucinations; weaknesses include limited task diversity and potential scalability concerns.

Read-first score

Read-first score 65, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 55.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

78.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,experiment,result

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 267.

Keyword Scores

world model

world dynamics prediction

world simulator

interactive world model

generative world model

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Using unstructured play data to learn an action-conditioned visual world model, diffusion-based action sampler, and optional reward model
Combining world model with Monte Carlo Tree Search (MCTS) planner and zeroth-order Model Predictive Controller (MPC) for planning
Action sampler mitigates hallucinations of the world model during planning

Methodology

The approach collects a few hours of unstructured easy-to-collect play data to learn an action-conditioned visual world model, a diffusion-based action sampler, and optionally a reward model. The world model, action sampler, and reward model are used to optimize long sequences of actions with Monte Carlo Tree Search (MCTS) planning, and the resulting plans are executed via a zeroth-order Model Predictive Controller (MPC). The method is validated on 3 real-world robotic tasks with varying levels of planning and modeling complexity.

Key Results

Planning leads to a significant improvement over Behavior Cloning (BC) baselines on a standard manipulation test environment.

Resources

Paper