Latency Budgets in Perception-Action Loops: Does Faster Always Mean Safer?

Control Frequency Against Stopping Clearance

The expectation behind a faster controller is straightforward: shorter update intervals should improve disturbance rejection and tracking. Check that expectation against the complete sensing–estimation–planning–actuation cascade. Trace each command to its originating observation, measure command acceptance, and compare requested torque with delivered response. Faster updates help where fresh information, scheduling headroom, and useful actuator bandwidth support them.

Define controller-side Age of Information as Δ(t) = t − u(t), where u(t) is the acquisition time of the newest physical observation incorporated into the estimate being consumed. Forward prediction advances the estimated state while the acquisition timestamp remains unchanged.

The clearance cost follows directly. At an illustrative closing speed of 12 m/s, an additional 40 ms of uncompensated observation-to-action delay consumes 0.48 m before braking travel enters the calculation.

Delay also changes stability margins. A pure delay τ contributes −360fτ degrees of phase at frequency f. An additional 10 ms therefore contributes −36 degrees at a 10 Hz gain crossover. Increasing scheduler frequency while retaining that delay leaves the phase penalty in place; raising feedback gains can move crossover into a less favorable part of the plant response.

Perception Queues and Observation Age

Put the queueing calculation first. For a stable M/M/1 queue, mean system time is W = 1/(μ − λ). With an illustrative service rate of 100 jobs/s, increasing arrivals from 50 to 90 jobs/s changes mean system time from 20 to 100 ms. Mean execution time remains 10 ms.

The processor completes more jobs while each observation spends longer inside the pipeline. Under the same assumptions, P(T > t) = exp[−(μ − λ)t]; shrinking service headroom stretches the latency tail as well as the mean. Throughput and freshness need separate measurements.

Use M/M/1 as an interpretable saturation baseline. Periodic camera acquisition with variable inference duration fits a D/G/1 description more closely; shared processors also require explicit treatment of scheduling and interference. Record acquisition, enqueue, execution-start, completion, and consumption timestamps so queue residence stays visible.

A controller repeatedly consuming the same estimate performs a hold operation. Evaluate its phase contribution together with sampling, transport, and prediction. The Age of Information in networked control and queueing systems perspective keeps the physical acquisition event central to that accounting.

Bound the Pending-Frame Queue

For a bounded, non-preemptive stage retaining only the newest pending frame, acquisition-to-completion age for the next pending completion is bounded by T_s + 2C_max + D_max. Here T_s bounds sensor intervals, C_max bounds execution, and D_max bounds additional transport. A FIFO backlog instead adds waiting associated with every queued frame. Include the wait for the first obstacle-containing observation, downstream processing, and actuator response when constructing the obstacle-reaction budget.

Actuator Bandwidth and Estimator Recovery

Queueing analysis leaves the command-to-motion response unresolved. Identify that response before choosing observer bandwidth or increasing tracking gains. Measure motor current-loop behavior separately from drivetrain compliance, hydraulic pressure dynamics, or series-elastic deflection.

A first-order actuator model, G_a(s) = 1/(1 + sτ_a), has bandwidth f_a = 1/(2πτ_a). An illustrative identified time constant of 20 ms gives approximately 8 Hz. Recomputing commands more frequently preserves that mechanical time constant.

Torque slew adds another constraint: Δt ≥ |Δτ|/r_τ, where r_τ is the admissible torque-change rate. Log requested torque, delivered current or estimated torque, and saturation flags on the same timeline. This exposes commands that the hardware spends successive updates trying to realize.

Recover the Estimate Before Raising Gains

Apply a controlled disturbance and inspect estimator innovation, covariance consistency, and recovery alongside actuator current. For a linear observer error mode with pole −α, decay to a fraction ε takes ln(1/ε)/α. Assess extended Kalman filter or moving-horizon recovery against a declared error or innovation threshold.

High-rate gains acting on noisy estimates can produce chattering, increased RMS current, thermal derating, and wear without corresponding center-of-mass acceleration. Measure winding-temperature estimates during rate sweeps.

Place Observer Modes

Use resonant peaks, the useful torque-to-motion band, and the sensor-noise spectrum to select observer modes and filter cutoffs. Recalculate discrete poles with z = exp(sT_s) whenever the update period changes, then evaluate the combined observer, filter, computation, and actuator phase at crossover.

Allocating the Acquisition-to-Clearance Budget

Build an acquisition-to-clearance ledger around a single physical observation. Assign capture and inference, estimator incorporation, replanning, and actuator response to the dependency graph. Obtain effective reaction time T_eff from its critical path; overlapping stages share elapsed time.

A useful screening model is d_required = v_cT_eff + v_c²/(2a_min) + e_perception + e_estimation + e_tracking. Positive closing speed v_c weights reaction travel; a_min is the minimum deceleration assumed for the operating envelope. The remaining terms allocate spatial uncertainty.

Charge actuator response through either an effective delay or the braking model. Treat estimator recovery as transient estimation error while control continues; count it as waiting only when the architecture gates action on convergence.

An asynchronous fast tracker can preserve rapid proprioceptive feedback while perception runs slowly. A coordinated multirate design can reserve compute headroom and bound scheduling jitter under declared execution assumptions. Compare them under the same sensing load, plant, and safety envelope: report acquisition-to-actuation age, deadline misses, command jitter, detection performance at the selected resolution, and tracking error. Asymptotic optimality of a planner addresses a separate property from these execution deadlines.

Profiling the Perception-to-Actuator Path

1. Reproduce deployed contention. Use hardware-in-the-loop operation with representative sensor traffic, message sizes, competing workloads, and an actuator interface. Preserve acquisition sequence identifiers across processes.
2. Separate timing components. Record execution duration, queue residence, inter-process transport, wake-up jitter, and command acceptance. Specify whether acquisition timestamps denote trigger, exposure midpoint, or scan timing; retain row or point timing where available.
3. Report the distribution. Keep mean and 99th-percentile latency separate, alongside sample count, maximum observed latency, dropped frames, and deadline misses. Sweep invocation rates while holding the task and controller configuration fixed.
4. Connect timing to phase. Plot the loop frequency response and overlay candidate delay contributions. At crossover ω_c, additional constant delay contributes −ω_cτ radians; the remaining-delay estimate is τ_margin = φ_m/ω_c with φ_m in radians. Replay measured variable-delay traces alongside this calculation.

For the frequency-domain construction, use the treatment of feedback systems and phase margin erosion under transport delay. Where tire slip or aerodynamic stall changes the plant during delayed feedback, move to protected hardware validation of the coupled dynamics before expanding the operating envelope.