Curiosity Based Exploration for Learning Terrain Models

Yogesh A. Girdhar, David Whitney, Gregory Dudek

Abstract

We present a robotic exploration technique in which the goal is to learn a visual model and be able to distinguish between different terrains and other visual components in an unknown environment. We use ROST, a realtime online spatiotemporal topic modeling framework to model these terrains using the observations made by the robot, and then use an information theoretic path planning technique to define the exploration path. We conduct experiments with aerial view and underwater datasets with millions of observations and varying path lengths, and find that paths that are biased towards locations with high topic perplexity produce better terrain models with high discriminative power, especially with paths of length close to the diameter of the world.

Introduction

Traditional robotic exploration usually focuses on spatial navigation—finding the boundaries between free space and obstacles to build a geometric map. However, in open environments like aerial surveying or underwater reef monitoring, obstacle avoidance is often secondary to semantic understanding. The challenge is twofold: low-level sensor data is often too noisy to represent abstract terrain concepts, and without a pre-existing map, a robot must decide where to move next to learn a generalized model as quickly as possible.

This paper introduces a "curiosity-driven" robotic exploration strategy designed to efficiently learn visual terrain models in unknown environments. By using a realtime spatiotemporal topic model to evaluate the semantic novelty (topic perplexity) of neighboring areas, the robot plans paths that maximize information gain. This approach significantly outperforms standard coverage or random walk strategies in building accurate terrain maps, especially when exploration time is strictly limited.

Methodology

The proposed system adapts topic modeling—a technique originally used in Natural Language Processing—for visual data. Instead of text, the system extracts "visual words" from images using ORB (for local features) and textons (for texture/color).

Realtime Online Spatiotemporal Topic Modeling (ROST)

The robot builds a topic model on the fly using ROST. To ensure topics cluster logically in physical space, the model uses Dirichlet priors that factor in the topic distribution of neighboring spatial cells. To maintain constant compute times as data grows, the algorithm employs Realtime Gibbs Sampling, updating only the current observation and one randomly sampled past observation per time step.

Curiosity-Based Routing

The robot decides its next move by calculating a weight for adjacent cells based on four potential strategies: Random Walk, Stochastic Coverage (repulsed by visited locations), Word Perplexity (drawn to novel low-level features), and Topic Perplexity (drawn to novel high-level semantic topics).

Overview of the experimental pipeline: (a-c) input maps, (d-f) ground truth segmentations, (g-i) exploration paths, and (j-l) resulting learned topic maps.

Experimental Setup

The algorithms were evaluated by simulating robot paths over three large maps: Montreal1 and Montreal2 (aerial satellite maps) and SouthBellairs (an underwater coral reef map). To ensure a diverse feature set, the visual word dictionary was pre-computed using features from the documentary film Baraka.

Results and Discussion

The experiments revealed that "Topic Perplexity" is a superior driver for exploration compared to low-level "Word Perplexity." Word perplexity often fails because it gets distracted by high-frequency visual noise. In contrast, topic perplexity abstracts this noise into semantic categories, allowing the robot to seek out genuinely new terrain types.

Performance Metrics

For short exploration paths (approximately the width of the maps), the Topic Perplexity strategy achieved significantly higher mutual information with human-labeled ground truth compared to other techniques. Specifically, it showed 1.51x, 1.20x, and 1.05x improvement across the three datasets.

Mutual Information between the learned model and ground truth as a function of path length, demonstrating the efficiency of curiosity-based exploration.

As path lengths increase, the Stochastic Coverage algorithm eventually matches the performance of Topic Perplexity, as the bounded nature of the maps ensures all terrains are eventually visited. However, curiosity-driven exploration remains the most efficient method when time is a constraint.

Conclusion

This work demonstrates that semantic novelty is a powerful heuristic for robotic exploration in unknown environments. By focusing on high-level topic perplexity rather than low-level visual complexity, robots can build accurate terrain models much faster than with traditional coverage or random strategies. Future work includes addressing dictionary domain shifts and extending the framework to 3D or dynamic environments.

Summary updated

Abstract

We present a robotic exploration technique in which the goal is to learn to a visual model and be able to distinguish between different terrains and other visual components in an unknown environment. We use ROST, a realtime online spatiotemporal topic modeling framework to model these terrains using the observations made by the robot, and then use an information theoretic path planning technique to define the exploration path. We conduct experiments with aerial view and underwater datasets with millions of observations and varying path lengths, and find that paths that are biased towards locations with high topic perplexity produce better terrain models with high discriminative power, especially with paths of length close to the diameter of the world.
BibTeX
@inproceedings{girdhar2014curiosity,
  author = {Yogesh A. Girdhar and David Whitney and Gregory Dudek},
  title = {{Curiosity Based Exploration for Learning Terrain Models}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2014},
  url = {https://openreview.net/forum?id=5hvvKRQ5BE},
}
Figure