Map Prediction and Generative Entropy for Multi-Agent Exploration

TL;DR

The paper addresses efficient multi-agent exploration under partial observability by predicting a complete map from partial observations using diffusion-based inpainting. It models per-cell occupancy with a belief $p_i=p(m_i)$ and defines generative entropy $H_i=-p_i\log_2 p_i -(1-p_i)\log_2(1-p_i)$ to guide task allocation, comparing diffusion-based inpainting methods (LaMa, RePaint, Stable Diffusion) on a procedurally generated urban dataset. Stable Diffusion is chosen for its speed and accuracy, and experiments show that prioritizing Generative Entropy accelerates convergence to high-quality maps, achieving a 62% faster high-accuracy convergence than a Visible Entropy baseline. The results suggest map prediction with generative-entropy guidance can substantially reduce exploration time and provide timely situational awareness in time-sensitive scenarios, with potential extensions to semantic conditioning and scene-graph reasoning.

Abstract

Traditionally, autonomous reconnaissance applications have acted on explicit sets of historical observations. Aided by recent breakthroughs in generative technologies, this work enables robot teams to act beyond what is currently known about the environment by inferring a distribution of reasonable interpretations of the scene. We developed a map predictor that inpaints the unknown space in a multi-agent 2D occupancy map during an exploration mission. From a comparison of several inpainting methods, we found that a fine-tuned latent diffusion inpainting model could provide rich and coherent interpretations of simulated urban environments with relatively little computation time. By iteratively inferring interpretations of the scene throughout an exploration run, we are able to identify areas that exhibit high uncertainty in the prediction, which we formalize with the concept of generative entropy. We prioritize tasks in regions of high generative entropy, hypothesizing that this will expedite convergence on an accurate predicted map of the scene. In our study we juxtapose this new paradigm of task ranking with the state of the art, which ranks regions to explore by those which maximize expected information recovery. We compare both of these methods in a simulated urban environment with three vehicles. Our results demonstrate that by using our new task ranking method, we can predict a correct scene significantly faster than with a traditional information-guided method.

Figures (4)

• Figure 1: Visualization of the generative entropy overlaid on the observed map during an exploration trial with three agents (paths shown in color). Dark regions indicate areas of high generative entropy, which are prioritized to resolve the remaining uncertainty in the predicted map. Task locations are shown as green squares.
• Figure 2: Map predictions for each model when 20%, 35%, 50%, 65%, and 80% of the map information was observed during a single trial. Masks of complex geometry (grey, second column) are inpainted by LaMa, Repaint, and Stable Diffusion. Performance is presented (last three columns) as correctly predicted free (white), correctly predicted occupied (black), incorrectly predicted free (green), and incorrectly predicted occupied (blue).
• Figure 3: Top row: simulated exploration of an urban environment with three agents, with images corresponding to 20%, 35%, and 50% cells observed. Bottom row: the corresponding generative entropy field. Darker pixels correspond to points with high entropy.
• Figure 4: Portion of map explored and accuracy of the predicted map over time, for three task reward settings. Each line shows the median performance across 10 trials, with the shaded area representing the interquartile range. Selected accuracy thresholds are indicated with dashed lines.