In classical control theory, engineers typically operate under an idealized assumption: sensors, controllers, and actuators communicate instantaneously, continuously, and with infinite precision. However, modern cyber-physical systems—such as robotic swarms, smart power grids, and distributed Internet of Things (IoT) networks—must operate over real-world digital channels.
As noted by Claire J. Tomlin et al., physical systems are fundamentally constrained by the digital infrastructure supporting them. Wireless channels possess finite bandwidth, suffer from random delays, and are powered by batteries that deplete rapidly under constant transmission. This reality introduces a core tension: physical stability demands high-frequency, high-precision updates, while the underlying communication network demands resource conservation.
To bridge this gap, the field of resource-constrained control shifts the paradigm from periodic operation (communicating at every tick of a fixed clock) to opportunistic operation (communicating only when mathematically necessary to preserve stability or performance).
The Core Trade-Off: Talk vs. Error
Every transmission over a shared network incurs a cost in bandwidth, energy, or latency. Conversely, withholding a transmission introduces state estimation error, as the controller must rely on stale data to calculate control inputs.
Erfan Nozari et al. formalized this mathematical trade-off, demonstrating a direct constraint between a system's exponential convergence rate and its communication frequency. If you want a system to stabilize faster, you must pay for it with more frequent updates. Shubham Aggarwal et al. framed this as balancing "the cost of talking with the cost of being wrong."
When multiple independent control loops share the same network, they must compete for these limited channels, as highlighted by Mohammad H. Mamduhi et al.. If every sensor samples at a high rate, the network suffers from congestion and packet collisions, which paradoxically degrades control performance (Carles Antón-Haro et al.; Dominik Baumann et al.).
Event-Triggered and Self-Triggered Execution
The most prominent strategy for minimizing communication is to replace periodic clocks with state-dependent execution. This comes in two primary flavors: event-triggered and self-triggered control.
Event-Triggered Control: Reacting to Deviation
In event-triggered control (ETC), a sensor continuously monitors the physical plant but only transmits data when a specific error threshold is crossed. For example, if the difference between the current state $x(t)$ and the last transmitted state $x(t_k)$ exceeds a designed boundary, an "event" is triggered, and a new packet is sent.
Burak Demirel et al. and Raffaello D'Andrea demonstrated that threshold-based ETC drastically reduces communication overhead while maintaining rigorous stability guarantees for unstable systems. The intuition is simple: if the system is behaving predictably or resting near its equilibrium, there is no reason to waste bandwidth saying so. This concept can also be extended to multi-agent consensus networks, where agents only "pull" updates from neighbors when their local tracking error warrants it (Xinlei Yi et al.).
Self-Triggered Control: Predicting the Next Update
While event-triggered control saves communication bandwidth, it still requires the sensor hardware to remain continuously awake to monitor the state. For battery-powered hardware, this continuous sensing is itself a massive energy drain.
Self-triggered control (STC) solves this by shifting the decision-making from the sensor to the controller. At the moment an update is received, the controller uses a mathematical model of the system's dynamics to calculate exactly how long the system can safely run open-loop before the state error risks violating safety or stability boundaries. It then schedules the next transmission time in advance, allowing the sensor's radio and processing units to enter a low-power sleep state in the interim (Kazumune Hashimoto et al.).
To make STC work for complex, nonlinear systems, researchers use hybrid Lyapunov functions. By introducing a dynamic variable that acts as a "memory" of past system performance, the controller can dynamically stretch the sleep intervals when the system is well-behaved (Michael Hertneck et al.; Michael Hertneck et al.). Even when transmission delays are present, these hybrid frameworks can maintain asymptotic stability while using a fraction of the samples required by traditional methods (Dynamic self-triggered control for nonlinear systems with delays). If the underlying physics of the plant are unknown, data-driven approaches like Gaussian Processes can be used to learn these safe sleep schedules on the fly (Learning self-triggered controllers with Gaussian processes).
Periodic: |-------|-------|-------|-------|-------| (Always communicate)
Event-Triggered: |-------|---------------|-------|----| (Communicate on error threshold)
Self-Triggered: |-------|-------------[sleep]---|-------| (Pre-calculate next wake-up time)
Information Limits and Quantization
Reducing how often we talk is only half the battle; we must also optimize how much we say during each transmission. Digital channels require continuous physical states to be converted into discrete bitstreams via quantization.
Victoria Kostina established fundamental Shannon limits for control, proving that instead of quantizing the raw state vector, a system should only quantize the innovation—the brand-new information that the controller could not have predicted using its internal model.
Furthermore, control performance exhibits diminishing returns as communication bit-rates increase. Dipankar Maity leveraged this insight by designing dynamic quantizer switching mechanisms. When the system is far from its target, it uses coarse, low-resolution quantization (saving bandwidth); as it approaches the target, it dynamically switches to high-precision quantization to eliminate steady-state error. This achieves near-perfect control with a remarkably low average bit-rate.
However, these information flows face strict physical ceilings. In adversarial environments or under channel blackouts, the resilience of a multi-agent system is strictly bounded by its weakest communication link (Pavankumar Tallapragada et al.). As Shuai Feng demonstrated, a deficiency in inter-agent communication capacity creates an unresolvable performance bottleneck that cannot be compensated for by increasing the data rate elsewhere.
Co-Design and Predictive Optimization
When scaling up to complex networks, communication constraints must be embedded directly into the control optimization problem itself. This is achieved through co-design, scheduling, and predictive control.
Actuator and Sensor Scheduling
If a network cannot support all actuators or sensors firing simultaneously, a scheduling mechanism must select the optimal subset for each time step. While finding the absolute best sequence is an NP-hard combinatorial problem, researchers use convex relaxation to make it tractable. By relaxing discrete selection variables into continuous weights, the problem can be solved efficiently in real-time, mapping the continuous solutions back to discrete actuator choices with guaranteed suboptimality bounds (Actuator Scheduling for Linear Systems: A Convex Relaxation Approach; Farhad Farokhi).
Predictive and Distributed Frameworks
Model Predictive Control (MPC) is uniquely suited for resource-constrained environments because it natively optimizes over a future time horizon while respecting constraints. By incorporating time-varying terminal regions and costs, MPC can optimize control actions and network transmission schedules simultaneously (Frank Allgöwer et al.). Under external disturbances, Tube MPC can be deployed to guarantee that rollout schedules provide updates frequently enough to keep the system within safe physical boundaries (Frank Allgöwer et al.).
In large-scale multi-agent systems, executing full optimization loops at every step can overwhelm both onboard computation and local wireless links (Gösta Stomberg et al.). To solve this, Giuseppe Belgioioso et al. proved that Distributed MPC does not need to reach full optimization convergence at each step. By maintaining a "running estimate" of the solution and communicating just a single round per sampling period, the network can maintain robust stability while drastically slashing its communication budget.
By systematically linking the physical dynamics of the plant to the digital constraints of the network, resource-constrained control ensures that modern autonomous systems remain safe, stable, and highly efficient.
Key Foundational Work
Numerical study on joint quantization and control under block-coding (2004) — contributed 1 concept elements to this direction.