RoboHorizon: An LLM-Assisted Multi-View World Model for Long-Horizon Robotic Manipulation

arXiv 25.1 2025 51.3 method, application

TLDR

RoboHorizon integrates LLMs and multi-view world models with keyframe discovery for long-horizon robotic manipulation, outperforming baselines on simulated benchmarks.

Reasoning

The paper presents a novel pipeline combining LLM-generated dense rewards and multi-view masked autoencoders for keyframe discovery, showing strong empirical gains on RLBench and FurnitureBench. However, it lacks real-world validation and the approach may be complex to deploy.

Read-first score

Read-first score 51.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.

Recency 8%

86.7

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

Methodology quality 25%

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

Topical relevance 42%

47.1

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

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 339.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

video world model

world simulator

generative world model

interactive world model

Deep Analysis

Innovations

Proposes the Recognize-Sense-Plan-Act (RSPA) pipeline for long-horizon robotic manipulation tasks
Introduces RoboHorizon, an LLM-assisted multi-view world model that leverages pre-trained LLMs to generate dense reward structures for multi-stage sub-tasks
Integrates keyframe discovery into multi-view masked autoencoder (MAE) architecture to enhance perception of critical task sequences
Combines dense rewards and multi-view representations to construct a robotic world model for efficient long-horizon planning and RL-based action

Methodology

RoboHorizon follows the Recognize-Sense-Plan-Act (RSPA) pipeline. Pre-trained LLMs generate dense reward structures for multi-stage sub-tasks based on task language instructions. Keyframe discovery is integrated into a multi-view masked autoencoder (MAE) to sense critical task sequences. A robotic world model is built using these dense rewards and multi-view representations, and RL algorithms are used for acting.

Key Results

On RLBench, RoboHorizon achieves a 23.35% improvement in task success rates on 4 short-horizon tasks and a 29.23% improvement on 6 long-horizon tasks from RLBench and 3 furniture assembly tasks from FurnitureBench, outperforming state-of-the-art visual model-based RL methods.

Resources

Paper