Separation Principles for Sensing and Control Co-design

This direction is merged into another direction: decoupled estimation and control. Its last version stays readable and citable here.

Version 1 of 19 papers18 conceptsfrozen 2026-10-04

In a perfect world, a robot would have an infinite array of sensors and actuators, providing a constant stream of high-fidelity data and precise movements. In reality, every sensor consumes power, every transmission eats up bandwidth, and every actuator adds weight. Engineers face a daunting combinatorial challenge: out of hundreds of possible sensor and actuator configurations, which subset provides the best performance for the lowest cost?

Traditionally, this "co-design" problem was seen as an intractable mess where sensing, estimation, and control were all tangled together. However, a series of breakthroughs has revealed a powerful separation principle that allows us to decouple these choices without losing optimality.

The Complexity of Choice

When resources are limited, you cannot simply turn everything on. You might be limited by a "budget"—perhaps you can only activate five sensors at a time to save battery, or you must choose a sparse set of actuators to minimize hardware complexity.

Mathematically, this is a selection problem. If you have $N$ sensors and can only use $k$, there are $\binom{N}{k}$ possible combinations. As $N$ grows, this number explodes, making exhaustive search impossible. Worse, it was long assumed that the "best" sensors to pick would depend intimately on the specific control law you intended to use, and vice versa. This coupling suggested that you had to solve for the sensors and the controller simultaneously.

A New Separation Principle

The core insight developed by Vasileios Tzoumas, Luca Carlone, and George J. Pappas is that for Linear-Quadratic-Gaussian (LQG) systems, this coupling is an illusion.

In Control and Sensing Co-design, the authors prove a separation theorem: the optimal control gains can be designed as if all sensors were available, and the optimal sensor subset can be chosen independently to minimize a specific "control-weighted" estimation error.

This is a massive simplification. It means the architecture of the system can be a sequential cascade:

  1. Select: Identify the sensor subset that minimizes the error most relevant to the control task.
  2. Estimate: Use a standard Kalman filter on those sensors.
  3. Control: Apply standard optimal control laws (like LQR).

This modularity, explored further in Sensing-Constrained LQG Control, ensures that the design of the "brain" (the controller) doesn't need to change just because the "eyes" (the sensors) are constrained.

Extending to Actuators and Scheduling

The logic of separation doesn't stop at sensors. Milad Siami and Ali Jadbabaie extended this framework to the joint scheduling of both sensors and actuators. In A separation theorem for joint sensor and actuator scheduling with guaranteed performance bounds, they demonstrate that the selection of active components can be decomposed into two independent sub-problems.

If you find a good approximation for the sensor schedule and a good approximation for the actuator schedule, their combination is guaranteed to be a good approximation for the entire closed-loop system. This allows designers to use "greedy" algorithms—picking the best next sensor one at a time—while maintaining rigorous bounds on how close the result is to the theoretical absolute optimum.

Why It Works: The Independence of Gains

The intuition behind why these systems decouple lies in the structure of the LQG problem. In standard LQG, the "Separation Principle" already states that estimation and control are independent. The research by Tzoumas and his colleagues pushes this further into the hardware domain.

Because the optimal control gain in an LQG setting depends only on the system dynamics and the cost weights—not on the noise covariance or the specific measurement matrix—the "best" way to move the robot doesn't change based on which sensors are active. The sensors only change how well we know where the robot is.

Even when measurements are "costly" (e.g., you pay a price every time you ping a sensor), the optimal strategy remains surprisingly structured. As shown in Infinite-Horizon Linear-Quadratic-Gaussian Control with Costly Measurements, the optimal sensing schedule often becomes periodic and independent of the system's state, allowing for predictable, low-overhead operation in large-scale networked systems.

From Theory to Scalable Algorithms

The ultimate value of these separation principles is computational. By proving that these problems can be decoupled, researchers have moved from NP-hard combinatorial explosions to efficient, polynomial-time algorithms.

As detailed in LQG Control and Sensing Co-Design, we can now provide provable performance guarantees for resource-constrained systems. We no longer have to guess if a sensor suite is "good enough"; we can quantify exactly how close a sparse, budget-friendly configuration is to the unconstrained ideal. This framework provides the mathematical foundation for the next generation of power-aware drones, bandwidth-limited IoT swarms, and cost-effective industrial automation.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_separation_principles_for_sensing_and_co_de0a5c00,
  title = {Separation Principles for Sensing and Control Co-design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/separation-principles-for-sensing-and-control-co-design/rdr_de0a5c003fd84b679cdf4254c45c0ae9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_separation_principles_for_sensing_and_co_de0a5c00_v1,
  title = {Separation Principles for Sensing and Control Co-design},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/separation-principles-for-sensing-and-control-co-design/rdr_de0a5c003fd84b679cdf4254c45c0ae9/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}