Testing

testing

The cross-cutting testing umbrella for robium. Robotics apps fail in layers a generic test suite misses — a node that never receives a message, a launch file that starts everything except the one node that mattered, a policy that looks fine on paper but never reaches the goal in sim. This skill frames the robotics-specific test pyramid and states the non-negotiable bar: a robium build is not done until its smoke test passes. It does not re-teach launch_testing/pytest mechanics for ROS 2 (that's ros2) or the lerobot-eval CLI (that's lerobot) — it frames both as test-pyramid layers and routes to them.

When to use this skill

• Setting up tests for any new robium project — this should be planned into the build from the start, not bolted on after the fact.
• The trigger phrases in the description: 'test the robot app', 'how do I test this node', 'smoke test', 'regression test in sim'.
• Before claiming any robotics app or sample works — a claim of "done" without a passing smoke test is not verified, it's a guess.
• Load alongside whatever skill is building the thing under test — ros2/ nav2/gazebo for the navigation vertical, lerobot for the manipulation vertical — this skill supplies the test framing, not a replacement for either.
• Cross-references — go to the sibling skill instead when the question is:
  • ROS 2 launch_testing/pytest mechanics themselves (fixtures, process actions, assertions) → the upstream launch_testing README and ROS 2 testing docs (linked in References). The ros2 skill does not yet carry launch_testing content — until it does, go upstream directly rather than hunting for it there. This skill frames launch/node testing as a pyramid layer either way.
  • The lerobot-eval CLI and its flags → lerobot. This skill frames policy eval as a test-pyramid layer with a pass/fail bar; lerobot owns the eval mechanics.
  • Setting up the simulator a regression test runs against → gazebo or isaac-sim (or simulation if the choice isn't made yet).
  • Where the test data comes from — worlds, robot models, sample datasets, fixture folders, goldens → test-assets. This skill decides what to test and when it passes; test-assets supplies what it runs against.
  • General (non-robotics) testing practices — unrelated to this skill; use whatever generic testing guidance already applies to the language/ framework in question.
  • The whole-stack decision this feeds into → architect (plans testing into the brief, doesn't bolt it on later).

Key directives

• Delegation posture: route + embed the pyramid framing. The test-pyramid structure, what belongs at each layer, and the trial-run bar live here; launch_testing/pytest syntax and lerobot-eval flags live downstream. Never re-teach either mechanism's syntax in this skill.
• A sample app is not done until its smoke test passes. This is the concrete acceptance bar for any robium trial run: "it builds" or "it starts without crashing" is not sufficient — a launch/node-level smoke test actually passing is the minimum evidence of "done." Treat a missing or failing smoke test as an open task, not a detail to fix later.
• Test at the right layer — don't push everything into sim. Logic that doesn't depend on ROS or a running robot belongs in a plain unit test, not a slow sim scenario; reserve sim-based regression tests for behavior that genuinely requires the simulated environment. See the pyramid in Decision guidance.
• Determinism is what makes a test a test. A sim scenario or policy eval that isn't seeded/repeatable produces flaky results that get blamed on the code instead of the test — apply simulation's determinism guidance (seeded physics, use_sim_time) and lerobot's small-scale-before-long-run discipline (a short, deterministic eval run as a CI gate, not a long training-scale rollout) to every automated test.
• Both verticals get equal treatment. The ROS/navigation vertical tests with launch_testing + pytest; the manipulation vertical tests with deterministic small-scale policy eval runs. Neither is "the real test framework" and the other an afterthought — plan both when a project spans both verticals.
• Assert against config constants, not re-typed literals. A regression test that hardcodes a rendered value (e.g. asserting the literal string --steps=2000) can keep "passing" while the fact it checks goes stale — a config constant dropped to 100 and the test never noticed. When a test asserts a value derived from a config constant, import and assert against the constant itself, not a re-typed copy (vla-trial).
• A measurement harness needs a guard that it didn't silently measure nothing. A benchmark CLI that catches and prints per-device exceptions can still exit 0 having measured zero devices — this happened on a transient HF 401 with no real code bug. Add a regression test for the harness itself, or a return 1 if measured == 0 inside it, so "ran clean" and "measured nothing" can't look the same (vla-trial).
• Gate every paid remote run behind a free local dry-run. Before submitting any paid remote job (a cloud GPU training run, a paid inference batch), run the pipeline-smoke layer end-to-end locally first — same pipeline, tiny scale, CPU is fine. It catches plumbing bugs (a camera-feature mismatch, a bad local-output-dir path) before they cost a paid run, not just before they cost wall-clock time (vla-trial).

Quick start

1. Identify what layer a new piece of behavior belongs at using the pyramid in Decision guidance — don't default to "write a sim test" for everything.

2. For ROS 2 apps: write plain pytest for ROS-independent logic, and a launch_testing case for anything that needs a running node or launch file — mechanics in the upstream launch_testing docs (see References). Run both through colcon test as part of the build.

3. For ML/policy apps: treat a short, deterministic lerobot-eval run against a fixed dataset/seed as the smoke test — mechanics in lerobot. Two distinct pass bars; don't mix them: a pipeline smoke (tiny train-from-scratch, then a few eval episodes) asserts exit codes and numeric metrics only — no success threshold, because an undertrained policy legitimately scores 0 (verified 2026-07-12, manip-trial); a regression gate (a handful of episodes against a known-good checkpoint) is where a stated success-rate threshold belongs. Save large-scale eval runs for manual validation, not every CI run — and don't let a benchmark masquerade as a regression test: a benchmark that re-derives an already-recorded number or loads a large model belongs marked slow/deselected from the default suite (e.g. pytest -m "not slow"), run on demand, not on every push. Leaving benchmarks in the default suite bloated one suite's runtime unnecessarily (3.5min → 8min, vla-trial).

4. Before declaring a sample app or feature done, run its smoke test and confirm it passes — this is the trial-run bar, not optional polish.

5. Wire the passing tests into CI so the bar holds on every change, not just the first time.

Decision guidance

The test pyramid for robotics (narrower and faster at the top, broader and slower at the bottom — most changes should be caught by a unit test, not a full sim run):

The trial-run bar: for any sample app or feature this pyramid covers, the minimum passing bar before calling it done is the node/launch smoke layer (ROS vertical) or a policy-eval smoke run (manipulation vertical) — sim-scenario and full-scale eval are the next layer up, expected for anything beyond a first working sample.

Platform gotchas

• CI runners are headless by default — sim-scenario tests must run headless (Gazebo server-only mode, Isaac Sim's runheadless.sh) rather than assuming a display; route the general headless strategy to environments if it isn't already decided.
• GPU-gated eval inherits its gate in CI too. A policy-eval smoke test that needs Isaac Sim inherits that skill's NVIDIA RTX GPU floor — confirm the CI runner actually has a qualifying GPU before wiring an Isaac-Sim-backed eval into every pipeline run, or keep CI on a gazebo/CPU-only eval path and reserve GPU-backed runs for manual/scheduled checks.
• macOS CI runners can't run the ROS 2 layers natively — the node/launch-smoke and sim-scenario layers need Docker on macOS, same as local development (see ros2's and gazebo's macOS gotchas); plan CI images accordingly rather than assuming a native macOS runner works.
• A free-space occupancy-map assertion must clear "unknown" gray, not just half-max (nav-trial, 2026-07-10). ROS trinary occupancy PGMs encode free=254, unknown=205, occupied=0, so a pixel >= 0.75*maxval free-space check counts unknown gray (205 > 191) as free — an all-unknown map then passes a free-space assertion falsely. Put the threshold above 0.80*maxval (0.9 works) so only genuinely-free pixels count.

Customization

• Small demo vs. a fleet-scale project: a one-robot sim demo may only need the top two pyramid layers (unit + smoke) to meet the trial-run bar; a production-bound project should build out sim-scenario and eval-threshold layers too — scale the pyramid's depth to the project's stakes, but never skip the smoke layer entirely.
• Adding a new vertical or module: extend the same four-layer pyramid rather than inventing a project-specific test taxonomy — the layer names and owning skills stay consistent across robium projects.

References

• Upstream: ROS 2 testing overview, ROS 2 launch package (launch_testing source), colcon test-result / test workflow, pytest documentation.
• Sibling skills: ros2 (launch_testing/pytest mechanics), lerobot (lerobot-eval mechanics and small-scale-run discipline), gazebo and isaac-sim (sim-scenario execution), simulation (determinism guidance this skill's sim-scenario layer relies on), test-assets (the data the pyramid's layers consume), environments (headless CI setup), architect (plans testing into the brief).

Changelog

• 1.4.0 (2026-07-31): nav-trial absorption — new Platform gotcha: a free-space occupancy-map assertion must threshold above 0.80maxval (0.9), because ROS trinary PGMs use free=254/unknown=205/occupied=0 and a 0.75maxval check counts unknown gray (205) as free, so an all-unknown map falsely passes.
• 1.3.0 (2026-07-18): route test-data sourcing to the new test-assets skill (cross-ref + sibling line); no content changes otherwise.
• 1.2.0 (2026-07-15): vla-trial absorption — Key directives gains three bullets (benchmarks aren't regression tests / mark slow; assert against config constants not literals; measurement harness zero-measured guard) and a free-local-gate-before-paid-remote-run rule; Quick start's large-scale-eval line extended with the slow-marker mechanism and 3.5→8min suite-bloat evidence.
• 1.1.0 (2026-07-12): manip-trial absorption — policy-eval layer split into pipeline-smoke (exit codes + numeric metrics, no success threshold) vs regression-gate (threshold against a known-good checkpoint); the previous universal-threshold phrasing made train-from-scratch smoke tests impossible to pass honestly.
• 1.0.1 (2026-07-11): nav-trial absorption — fixed the launch_testing routing dead-end: three spots routed "mechanics → ros2" but the ros2 skill has no launch_testing content; routes now point at the upstream docs until ros2 grows that section. Trial-run bar + smoke shape confirmed ✓ under real load (one-command compose smoke, exit-code chain).