When a robot enters an unknown environment, its primary goal is to build a map. Traditionally, this is done using "frontier exploration"—simply moving toward the boundary between known and unknown space. However, this geometric approach treats all unknown areas as equally valuable. In reality, some movements provide far more information than others.
The research direction of Information-Theoretic Exploration shifts the focus from "where has the robot not been?" to "where can the robot learn the most?" By treating the environment as a probabilistic belief state, robots can use the mathematical concept of entropy to quantify their own ignorance and plan paths that maximize Information Gain.
From Occupancy to Uncertainty
In a standard grid map, a cell is either "free" or "occupied." But sensors are noisy, and environments are complex. Cyrill Stachniss and Wolfram Burgard moved beyond this binary view in Exploring Unknown Environments with Mobile Robots using Coverage Maps, introducing maps that represent the degree of occupancy.
This probabilistic framing allows us to measure the entropy of the map—a single number representing how much we don't know. A cell with a 50% chance of being occupied has maximum entropy (total uncertainty), while a cell with a 0% or 100% chance has zero entropy (total certainty). Exploration then becomes a process of systematic entropy reduction.
Planning in Belief Space
If the goal is to reduce entropy, the robot must evaluate potential actions based on their expected outcome. This is known as Belief Space Planning. Instead of planning a path through physical coordinates $(x, y)$, the robot plans through the space of its own probability distributions.
As Atanas Mirchev and colleagues describe in Approximate Bayesian Inference in Spatial Environments, the agent selects actions that maximize the mutual information between its future observations and its internal map. This mechanism naturally incentivizes "curiosity": the robot is drawn to areas of high mathematical uncertainty without needing hand-coded rewards.
This approach is particularly powerful when using continuous representations. Thai Duong and Nikolay Atanasov demonstrated in Autonomous Navigation in Unknown Environments with Sparse Bayesian Kernel-based Occupancy Mapping that by using Bayesian models, robots can identify high-variance regions and plan paths that "smooth out" the map's uncertainty more efficiently than grid-based methods.
Infotaxis: Searching Without a Gradient
Information-theoretic strategies are essential when a robot is searching for a specific target, like a gas leak or an underwater object, where signals are sparse or "patchy." In these environments, there is no clear gradient to follow.
The strategy of Infotaxis, explored by Christopher Gilliam in Decentralised Scalable Search for a Hazardous Source in Turbulent Conditions, directs agents to move in ways that minimize the entropy of the target's predicted location. Rather than moving toward the strongest signal (which might be a fluke), the agent moves to where a measurement—even a "null" measurement—would most effectively narrow down the search area. Alex Skvortsov further refined this for underwater targets in Intermittent Information-Driven Search for Underwater Targets, showing that alternating between fast travel and slow, information-rich sensing is the optimal way to find targets in the shortest time.
Active Perception and Semantic Clarity
Modern research has extended these ideas from simple geometry to complex semantics and 3D reconstruction. It is no longer enough to know where an object is; the robot needs to know what it is.
In Understanding while Exploring: Semantics-driven Active Mapping, Liyan Chen and colleagues use semantic entropy to guide robots. If a robot sees a blurry shape that could be either a chair or a table, the high semantic entropy at that location triggers the robot to move for a better vantage point. Similarly, Yanyan Li and Gim Hee Lee developed Active3D, which uses hierarchical uncertainty to plan camera paths for high-fidelity 3D reconstruction, ensuring that complex surfaces are refined while simple ones are merely noted.
Why It Works: The Efficiency of Information
The core intuition behind this research is that information is a more efficient objective than distance.
Traditional path planning minimizes travel cost, but Thomas Kollar and Nicholas Roy showed in Efficient Optimization of Information-Theoretic Exploration in SLAM that prioritizing information gain can actually lead to shorter total exploration times. By revisiting known areas to reduce localization drift or taking a slightly longer path that provides a clearer view of a room, the robot avoids the "re-work" required when a map becomes too noisy to use.
Whether it is a UAV swarm mapping weeds in a field (Carlos Carbone et al., Monitoring and mapping of crop fields with UAV swarms based on information gain) or a lunar rover searching for ice in the dark (Yuki Mori, Modeling of observation instruments and path planning...), the principle remains the same: the most intelligent path is the one that resolves the most doubt.