DREAMer-VXS: A Latent World Model for Sample-Efficient AGV Exploration in Stochastic, Unobserved Environments

arXiv 25.10 2025 64 method

TLDR

A model-based RL framework using a latent world model (VAE+RSSM) for sample-efficient AGV exploration, achieving 90% fewer interactions than SAC.

Reasoning

The paper presents a clear methodology and strong empirical results in simulation, but lacks real-world validation and does not address video-based world modeling. The core contribution is the world model for planning, which is well-supported by the abstract.

Read-first score

Read-first score 64, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 50.

Recency 8%

86.7

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

Topical relevance 42%

71.4

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

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

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 180.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

DREAMer-VXS: a model-based framework for AGV exploration using a latent world model learned from partial LiDAR observations.
Integration of a Convolutional VAE and Recurrent State-Space Model (RSSM) to learn compact representations and temporal dynamics from high-dimensional sensory data.
Decoupling policy learning from real-world interaction by training entirely in imagination, achieving a 90% reduction in required environmental interactions.
Composite reward function combining task objectives with an intrinsic curiosity bonus to promote systematic exploration in stochastic, unobserved environments.

Methodology

DREAMer-VXS learns a world model consisting of a Convolutional Variational Autoencoder (VAE) for compact state representation and a Recurrent State-Space Model (RSSM) for modeling temporal dynamics from partial LiDAR observations. The agent uses this learned model as a high-speed simulator to train an actor-critic policy entirely in imagination, guided by a composite reward that balances task goals with an intrinsic curiosity bonus. The framework is evaluated in simulated environments against a state-of-the-art model-free SAC baseline.

Key Results

DREAMer-VXS achieves a 90% reduction in environmental interactions to reach expert-level performance compared to SAC, and demonstrates a 45% increase in exploration efficiency in unseen environments with superior resilience to dynamic obstacles.

Limitations

Only validated in simulated environments; real-world applicability remains unverified.
Relies on partial LiDAR observations which may not capture full environmental state, potentially limiting generalization.
The composite reward function with curiosity bonus may require careful tuning for different tasks or environments, and its effectiveness in highly stochastic settings is not fully characterized.

Resources

Paper