🧬 biology

Active Sensing Subserves Task-Level Control

This paper proposes that active sensing is not primarily driven by sensory goals like uncertainty reduction, but rather emerges as a necessary strategy for robust task-level control, where organisms switch between discrete "explore" and "exploit" modes to shape sensory feedback—a biological principle that could significantly advance the design of engineered robotic systems.

Original authors: Andrew Lamperski, Debojyoti Biswas, Eric S. Fortune, John Guckenheimer, Kathleen Hoffman, Noah J. Cowan

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrew Lamperski, Debojyoti Biswas, Eric S. Fortune, John Guckenheimer, Kathleen Hoffman, Noah J. Cowan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to balance a broomstick on your palm. To keep it upright, you have to move your hand constantly. But here's the twist: your hand isn't just moving to stop the broom from falling; it's also moving to feel the broom better.

This paper argues that in the animal kingdom, "active sensing" (moving to gather information) isn't just a luxury for learning about the world. Instead, it is a necessary survival trick that happens automatically because of how our bodies and brains are built.

Here is the breakdown of their discovery, using simple analogies:

1. The Problem: The "Numb" Sensor

Most biological sensors (like our eyes, ears, or the electric sense of a fish) are adaptive. This means they are great at noticing changes, but they get "numb" to things that stay the same.

The Analogy: Think of walking into a room with a strong smell of coffee. At first, you smell it intensely. Ten seconds later, your nose "adapts," and you stop noticing it, even though the coffee is still there.
The Consequence: If an animal stays perfectly still, its sensors go "blind" to its own position. It can't tell where it is relative to the world. To stay in control, it must move to create new changes that its sensors can detect.

2. The Solution: The "Explore" and "Exploit" Dance

The paper suggests animals don't just move randomly. They switch between two distinct modes, like a driver changing gears:

The "Exploit" Mode (The Cruise Control):
- Goal: Get the job done efficiently (e.g., stay in a moving boat, track a prey).
- Action: The animal moves slowly and smoothly to cancel out errors. It tries to stay still relative to its target.
- Analogy: You are driving on a straight highway, hands steady, just trying to stay in your lane. You aren't looking around much because you know where you are.
The "Explore" Mode (The Check-Up):
- Goal: Wake up the sensors.
- Action: The animal makes quick, jerky, or larger movements. This creates "sensory slip" (a mismatch between where it is and what it feels), which jolts the sensors back to life.
- Analogy: You are driving in fog. You can't see well, so you tap the brakes or swerve slightly to see how the car reacts and to get a better read on the road. You are moving specifically to get information.

The Magic: Animals switch between these two modes incredibly fast. When their internal "uncertainty meter" gets too high (they aren't sure where they are), they switch to Explore to get a fix. Once they know where they are, they switch back to Exploit to finish the task.

3. Why Nature Does This (But Robots Don't)

Engineers usually build robots with "perfect" sensors that don't get numb. They use a rule called the Separation Principle:

First, the robot calculates exactly where it is.
Then, it decides how to move.

Nature doesn't do this. Because biological sensors are "adaptive" (they get numb), the robot cannot calculate where it is without moving first.

The Paper's Claim: Active sensing isn't a separate step to "learn" about the world. It is an unavoidable side effect of trying to control a body with adaptive sensors. You have to wiggle to know where you are, and you wiggle because you need to control your movement.

4. The "Weakly Electric Fish" Example

The researchers studied a small fish called Eigenmannia that swims inside a moving tube (a refuge).

The fish has an electric sense that gets "numb" if the water is still.
To stay in the tube, the fish has to constantly wiggle back and forth.
When the water is dark (hard to see), the fish wiggles more (more "Explore" mode) to make sure it doesn't drift out of the tube.
This proves that the movement isn't just for "looking"; it's mathematically required to keep the fish stable.

5. The Big Takeaway

The paper argues that we have been looking at animal movement the wrong way. We thought animals moved to "gather information."
The new view: Animals move to control their task, and gathering information is just a necessary byproduct of that movement.

Old View: "I am moving my head to see the ball better."
New View: "I am moving my hand to catch the ball, and my brain is forced to wiggle my head to keep my sensors from going numb so I can actually control my hand."

The authors suggest that if we want to build better robots, we shouldn't just give them better cameras. We should teach them to "wiggle" (switch between exploring and exploiting) just like animals do, because that is how you stay in control when your sensors aren't perfect.

Technical Summary: Active Sensing Subserves Task-Level Control

Problem Statement
Traditional definitions of active sensing describe it as the expenditure of energy (typically movement) to achieve sensory goals, such as minimizing uncertainty about an organism's state or reducing sensory prediction error. However, this perspective assumes that information gathering is the primary driver of movement. The authors identify a gap in understanding: how active sensing movements emerge in biological systems that rely on adaptive sensors (sensors that reject constant baselines and respond primarily to changes in stimuli). Adaptive sensing, while advantageous for sensitivity across wide dynamic ranges, creates a fundamental observability problem. In dynamical systems terms, adaptive sensors render the system locally unobservable at equilibrium, theoretically precluding stable closed-loop control of task-oriented behaviors if the system remains static. The paper questions whether active sensing is driven by "information goals" or if it is an inescapable consequence of the physical and control-theoretic constraints imposed by task-level feedback control.

Methodology
The authors employ a multi-faceted approach combining mathematical theory, computational modeling, and empirical analysis of biological data:

Theoretical Modeling: They utilize a simplified second-order template model of the weakly electric fish (Eigenmannia virescens) locomotor plant. This model incorporates an adaptive sensor that responds to the time derivative of the sensory stimulus ( $y = \gamma(x_1)x_2$ ). The authors demonstrate mathematically that for such a system, stabilization to an equilibrium point is impossible without active sensing movements. Stability can only be achieved around time-varying movements (limit cycles).
Computational Simulations: The authors simulate control strategies comparing:
- Persistent Excitation (PE): A classical control-theoretic strategy where rich, varying inputs are continuously applied to ensure parameter identifiability.
- Mode-Switching (Explore/Exploit): A heuristic strategy where the controller switches between an "exploit" mode (minimizing tracking error with slow, compensatory movements) and an "explore" mode (generating fast, active sensing movements to reduce state uncertainty).
- These simulations utilize a non-convex optimization landscape where the cost function includes tracking error and control effort, but not an explicit term for information gain.
Empirical Analysis: The authors re-analyze time-series data from a diverse set of organisms (ranging from nematodes to humans) performing station-keeping tasks. They quantify behavioral modes and switching dynamics, specifically looking at how sensory salience (e.g., light levels for vision, or jamming for electrosense) influences the frequency of transitions between modes.

Key Contributions

Reframing the Origin of Active Sensing: The paper proposes that active sensing is not necessarily driven by a sensory goal (like minimizing uncertainty) but emerges inevitably from the requirements of task-level feedback control when the system relies on adaptive sensors. The combination of goal-directed behavior, closed-loop control, and geometric/dynamic sensory constraints forces the generation of self-movements that do not directly achieve the task goal but enable the task to be performed.
The Explore-Exploit Mode Switching Framework: The authors characterize animal behavior as a rapid switching between two distinct control policies:
- Exploit Mode: Slow, compensatory movements designed to minimize sensory slip (error) and achieve the task goal directly.
- Explore Mode: Fast, active sensing movements designed to increase sensory slip and improve state estimation.
- This switching is triggered by internal estimates of uncertainty; when uncertainty exceeds a threshold, the system switches to explore mode to reduce it.
Superiority of Mode Switching over Persistent Excitation: Theoretical results show that the mode-switching strategy outperforms classical persistent excitation methods. In simulations, mode-switching achieves lower tracking error and reduced control effort (energy cost) compared to continuous excitation, particularly in non-convex optimization landscapes where the cost function does not explicitly reward information gathering.
Critique of the Separation Principle: The paper argues that the engineering "separation principle" (estimating state separately from control) fails for systems with adaptive sensors. Unlike engineered systems that use absolute, non-adaptive sensors, biological systems must integrate sensing and actuation, making the generation of active sensing movements a necessary part of the control policy itself.

Results

Mathematical Necessity: The analysis of the adaptive sensor model proves that without active sensing movements (specifically, movements that create a time-varying stimulus), the system is unobservable and cannot be stabilized to a fixed point.
Behavioral Evidence: Empirical data confirms that animals across diverse taxa (including glass knifefish, cockroaches, and humans) exhibit discrete epochs of fast "explore" movements interspersed with slow "exploit" movements.
Sensory Salience Effects: Reduced sensory salience (e.g., darkness or electrosensory jamming) leads to a significant increase in the frequency of transitions to the "explore" mode and a higher proportion of time spent in high-velocity active sensing, consistent with the hypothesis that uncertainty drives the switching.
Performance Gains: Simulations demonstrate that uncertainty-triggered switching strategies achieve better performance (lower cost) than continuous excitation strategies, even when the cost function only penalizes tracking error and control effort, not information acquisition.

Significance and Claims
The paper claims that the current understanding of active sensing is incomplete because it focuses on information goals. Instead, the authors argue that active sensing is a necessary consequence of control in systems with adaptive sensors. This insight suggests that the robustness and grace of animal behavior, which often outperform engineered systems in terms of robustness and adaptability despite inferior hardware, stem from this specific control architecture.

The authors assert that engineering systems, which typically rely on the separation principle and non-adaptive sensors, miss a critical component of biological intelligence. By adopting adaptive sensors, active sensing, and mode-switching control policies, robotic systems could potentially achieve similar levels of robustness and efficiency. The paper positions these findings not as a new optimization algorithm for information gathering, but as a fundamental shift in how control systems for embodied agents should be designed: acknowledging that the "plant" (the body and its sensors) and the controller are inextricably linked, and that movement is sometimes required not to change the state, but to make the state observable.