Hardware-in-the-Loop

Relevant source files

There is no hardware-in-the-loop CI. Raspberry Pi 5 testing is manual (README, issue 76). Nothing on this page runs in GitHub Actions or in cargo xtask check all. At this release the repository contains no HIL runbook and no scripts/hil/ directory; the tooling below is what exists.

docs/operations/release-verification.md states that physical HIL and independent review remain separate evidence and must not be inferred from a local gate pass. The v0.5.0-alpha.1 scope in that file explicitly does not claim physical robot HIL, hosted CI or a ForgeFPGA bitstream. The extended profile in ci/profiles/extended.toml lists an hil check id, but engineering-audit.sh contains no physical-hardware step.

Build and deploy

Pi 5 images are built and deployed with xtask; see Raspberry Pi 5 for detail.


export AXIOM_BPF_TRUSTED_KEY_PATH=/secure/path/axiomos-ed25519.pub
cargo xtask build rpi5 -- release embedded-rpi5
sha256sum -c target/aarch64-unknown-none/release/rpi5-artifacts.sha256
cargo xtask deploy rpi5 -- /path/to/mounted/pi-boot

The `bench` feature

kernel/Cargo.toml defines a bench feature for the v0.3 hardware bench: on-chip latency instrumentation (M-A, M-B, M-C), the physical e-stop button IRQ path, and an auto-loaded GPIO-to-stop reflex. It is off by default and is measurement and demo instrumentation, not for production images. Enable it alongside a profile:


cargo xtask build rpi5 -- release embedded-rpi5,bench

Metric	Meaning
M-A	Actuation monitor overhead (`[bench] M-A monitor_overhead_ns=N`)
M-B	E-stop IRQ entry to safe state (`[bench] M-B estop irq-entry->safe latency_ns=N`)
M-C	Edge to PWM or GPIO apply (`[bench] M-C edge->pwm-apply ch=C val=V latency_ns=N`)

Analyzing captures

scripts/benchmark/analyze-v03.py parses [bench] serial lines and, optionally, a logic-analyzer CSV with columns for the GPIO23 input, GPIO12 PWM enable and GPIO24 e-stop (common aliases such as ch0, ch1, ch2 are accepted).


scripts/benchmark/analyze-v03.py --serial bench.log [--logic capture.csv]
scripts/benchmark/analyze-v03.py --self-test

It prints count, min, median, p99, p99.9 and max per metric and fails when thresholds are missed:

Check	Fails when
M-A	No samples, or max is 5000 ns or more
M-B	No samples, or max is 1 ms or more
M-C	Fewer than 10,000 samples, or median is 1000 ns or more (the 1000 ns fallback; a median from 500 to under 1000 ns prints that the 500 ns target was missed)

The --self-test run is part of the command-smoke step in the audit gate. The gate does not run any physical campaign.

Verifier cost on the Pi 5

A kernel built with --features verifier-cost emits an AXIOM VERIFIER COST marker per BPF load. scripts/benchmark/verifier-cost.py (cargo xtask bench verifier) turns the UART log into a CSV and plot; see Benchmarks.

Evidence

Hardware results follow the same contract as any other campaign: exact commit, pinned toolchain, command, hardware boundary, raw capture, and hashes of the raw log and measured image. See the evidence policy. At this release the only attributable campaign is the host BPF verifier benchmark; see Current Results.