The Core Dilemma: Information is Not Free
Imagine a mobile robot tasked with finding a specific object in a large, unfamiliar warehouse. A passive robot sits still, processing whatever data happens to hit its sensors. An active robot, by contrast, intentionally changes its viewpoint or interacts with its environment to resolve ambiguity—such as moving around a pillar to see what is hidden behind it.
In theory, an active perception agent should always seek to maximize its information gain or minimize its internal uncertainty (entropy). However, as Cohn points out, physical data collection is inherently resource-intensive, incurring significant costs in time, energy, and mechanical wear. If a robot travels a long distance just to resolve a tiny fraction of uncertainty, the energy spent outweighs the value of the information gained. This creates a fundamental tension between the informational utility of an action and its physical execution cost.
Mathematically Balancing the Scales
To resolve this conflict, researchers have shifted from pure information maximization to cost-aware optimization frameworks. The core idea is to modify the robot's objective function so that it explicitly subtracts physical costs from expected informational rewards:
$$\text{Utility} = \text{Expected Information Gain} - \lambda \cdot \text{Physical Cost}$$
Here, $\text{Physical Cost}$ can represent travel distance, time, or energy expenditure, while $\lambda$ acts as a scaling factor balancing the two objectives.
This formulation fundamentally changes how a robot plans its actions. For example, Burgard et al. incorporated a cost function into active localization, using value iteration to ensure that a robot only pursues distant landmarks when the expected reduction in uncertainty justifies the navigation cost. Similarly, Taniguchi et al. demonstrated that subtracting a weighted travel distance from expected information gain prevents robots from over-prioritizing distant, high-uncertainty map boundaries. This trade-off results in more efficient exploration of central, functional areas, producing movement patterns that closely resemble human navigation.
Application Across Modalities
The information-cost tradeoff manifests across three major domains of robotics:
1. Autonomous Exploration and Mapping
When mapping unknown environments, a robot faces a choice between visiting a nearby, partially certain area or a distant, completely unknown frontier. Stachniss et al. showed that while prioritizing high-information vantage points reduces the total number of measurements needed, it can drastically increase path length. To mitigate this, Liu et al. utilized a composite cost function evaluating targets based on path length, sensor range, and information gain. In path planning under uncertainty, Zheng noted that a longer path through an information-rich area might be safer for a robot with noisy sensors, meaning the algorithm must carefully calculate when uncertainty reduction justifies the extra movement cost.
2. Object Detection and Target Search
In visual search, exhaustive exploration is often prohibitively expensive. Vélez et al. integrated motion costs directly into perception planning, preventing robots from making inefficient detours for marginal gains in detection confidence. By explicitly reasoning about occlusions, Li et al. showed that a robot can confirm an object's existence with roughly 90% fewer steps than an exhaustive search. For aerial platforms like UAVs, Jiang et al. designed reward functions that penalize long flight paths, encouraging the drone to settle on the first sufficient viewpoint rather than searching indefinitely for a perfect one.
3. Interactive Manipulation
When a robot must learn physical object properties (like mass or stiffness), it must physically interact with the environment. Amiri et al. and Wu et al. leveraged uncertainty-driven exploration strategies to minimize the number of physical pushes or shakes required to identify objects in dense clutter, directly reducing resource expenditure. Kruzliak et al. emphasized that because physical interaction causes mechanical wear, prioritizing the most informative interactions is a necessity for practical deployment.
Why It Works: Diminishing Returns and Regularization
Why does explicitly penalizing cost work so well? The answer lies in two phenomena: diminishing returns and behavioral regularization.
First, information acquisition in a physical environment obeys a law of diminishing returns. As Das et al. observed in autonomous environmental sampling, once a mobile unit has sampled high-variance regions, spending additional energy to gather more data provides negligible improvements to the overall model error. A cost-blind agent will continue to expend energy for these marginal gains, whereas a cost-aware agent will stop once the information profit turns negative.
Second, explicitly penalizing action costs acts as a powerful regularizer for behavior. In deep reinforcement learning, agents trained purely to maximize information or task rewards often develop highly unnatural, jittery behaviors—such as constant shaking or spinning—to maximize sensor coverage. Ho et al. demonstrated that introducing adaptive behavioral costs tames these non-naturalistic behaviors. By forcing the agent to pay a price for rapid or repetitive movements, the system naturally balances efficient information acquisition with smooth, energy-efficient physical trajectories.
Key Foundational Work
Active Mobile Robot Localization (1997) — contributed 3 concept elements to this direction.
Neural Network Exploration Using Optimal Experiment Design (1993) — contributed 1 concept elements to this direction.