DAPPER

Deadline-aware perception placement

    DAPPER
    booting — 0 parity —

    Deadline-aware edge perception · IEEE FMEC 2026

    watch it decide

    An answer after the deadline isn't a late answer. It's no answer.

    Mission-critical robots have milliseconds. A stronger model on the edge is worth having, but it puts the network in the critical path, and the right place to run a frame changes frame by frame. DAPPER scores each frame's deadline risk, places it, and refuses any remote result that arrives outside its freshness window.

    Open the console Read the paper

    Synthetic network benchmark. Confidence values are simulated proxies, never detector accuracy, and nothing here ran on a robot.

    dapper · frame 0 live

    —

    Loading the frozen configuration and a published scenario trace.

      deciding gate

      —

      — —

      Live · this is the real scheduler, not a mock

      The console

      Every frame takes the path the execution model recorded.

      A local inference that never leaves the robot. A hybrid request that answers locally now and asks for a refinement anyway. A lost packet. A reply rejected at the deadline. A degraded-safe reuse of the last still-fresh output. The ring closes on the frame's real control latency.

      frame 0 Rt — RTT — control latency — —
      placement local-fast edge-accurate (commits the frame) hybrid (local answer + refinement) degraded-safe (reuse or local)
      remote reply accepted, on time and fresh rejected at the deadline lost, or the edge was unreachable

      How it decides

      One risk score, two thresholds, three gates — and a reason string for every frame.

      This is the paper's Fig. 1, lit by the reason the scheduler actually emitted. Underneath it, the numeric test behind each gate, evaluated on the frame you are looking at.

      Decision path

      —

        Deadline risk Rt

        five weighted terms
        — —

        Telemetry

        What the monitor sees, and what the control loop gets.

        Network state RTT (ms), loss, load
        Control latency vs deadline —
        Usable confidence 0 when late or stale
        Mode share so far local / edge / hybrid / degraded

        Drive it yourself

        Move the deadline and watch the calibration stop holding.

        Parameters were frozen at a 100 ms deadline. Drag D past it and the committing edge-accurate branch starts taking frames — and starts missing.

        —

        —
        —
        50–250 ms. The paper calibrates at 100 ms only.
        —
        Perturbs only what the scheduler believes.
        Negative = underestimate, the dangerous direction.
        Marks the edge unreachable from the next frame.
        Switch one off to reproduce a row of Table V.

        Policy race

        Five policies, one immutable future.

        Each replays the identical trace through the identical cost model, so a decision selects which outcome is realised but cannot change the random future. The oracle reads the realised outcome of the current frame and is not deployable; it is here to bound what perfect knowledge of the next instant would buy.

        —

        Published results

        Every number here was read from a CSV by a script.

        —

        p95 control latency by profile Fig. 2a, 95% CI over 30 seeds
        Usable confidence by profile Fig. 2b
        Mode share vs deadline Fig. 3a, pooled
        Misses vs estimate bias Fig. 3b, profile mean

        The paper

        DAPPER: Deadline-Aware Edge Perception for Mission-Critical Robots under Dynamic Network Conditions

        Awaiz Ahmed and Khubaib Amjad Alam · Al Ain University, Abu Dhabi, United Arab Emirates · IEEE FMEC 2026

        Read online PDF Artifact

        Where DAPPER does not win