IROS Experiments
Use this before submission when the empirical story is not yet locked. At IROS, experiments exist to
prove a robot did something reliably, under stated conditions — not to top a leaderboard.
Experiment audit
- Map each claim to the evidence that supports it: a real-robot trial set, a simulation study, an
ablation, or a field deployment. A claim with no matched evidence is an overclaim.
- Define the success criterion before running, in one sentence a skeptic would accept, and hold to
it; a criterion invented after seeing results is a reviewer red flag.
- Report trial counts and resets: how many attempts, how start conditions were randomized, and what
reset happened between trials. Hidden resets inflate reliability.
- Publish a failure taxonomy with counts, not just a success rate; the failures are where reviewers
calibrate trust.
- Make baselines fair: run them on the same platform, sensors, and route, with comparable tuning
effort, or state precisely why not.
- Separate simulation from real and report the sim-to-real gap as a number; an implied zero gap
is the fastest way to lose a reviewer.
The evidence ladder
| Claim altitude |
Evidence IROS expects |
Reject pattern avoided |
| "The system works" |
Trials with n, success interval, and resets stated |
"One hero run shown as if typical" |
| "It is reliable" |
Failure taxonomy with counts across conditions |
"Success rate with no failures reported" |
| "It transfers" |
Real-robot numbers plus the measured sim-to-real gap |
"Sim results implying real performance" |
| "It beats prior work" |
Same-hardware baseline on the same task |
"Comparison against a weaker or re-tuned baseline" |
| "It runs onboard" |
Measured rate and power under the real compute budget |
"Real-time asserted, never measured" |
Small-n statistics for real robots
Real trials are expensive, so n is small — but small n does not excuse a bare mean. Report a success
count as a proportion with a confidence interval (a Wilson interval behaves better than normal
approximation at small n), and for paired system-vs-baseline comparisons on the same trials, prefer a
paired test over independent means. State n every time; "usually succeeds" is not a measurement.
Vignette: a manipulation reliability study
Suppose the system claims reliable grasping of unseen objects. The matching plan: fix an object set and
a success criterion (lifted and held 3 seconds), run a stated number of trials per object with
randomized poses, log every failure by cause (slip, mis-localization, collision), and report the
per-object and pooled success rates with intervals. A simulation sweep over object mass then maps where
the grasp model degrades, and the real-vs-sim gap is stated — every panel tied to a specific claim.
Trial-logging template (one row per attempt):
trial_id, object/scenario, start_pose_seed, outcome{success|fail},
failure_cause, reset_type, wall_time, notes
Aggregate to: success rate + interval per condition, failure histogram, sim-to-real delta.
Reporting floor
- Every reliability figure carries n and a criterion; every timing claim carries a measured rate and the
compute/power budget it ran under.
- Report the compute actually consumed on the robot, not a desktop proxy.
Output format
[Experiment readiness] strong / adequate / weak
[Claim -> evidence map] <claim: trials/sim/ablation/deployment>
[Missing evidence] <trials/resets/failures/baseline/transfer>
[Statistics] criterion set? interval reported? paired where paired?
[Decision-critical next run] <one experiment on the robot>
1---2name: iros-experiments3description: Use when designing or auditing IROS experiments — real-robot trial counts, success criteria set in advance, reset procedures, failure taxonomies, baseline fairness on matched hardware, sim-to-real gap reporting, small-n statistics, and the claim-to-evidence ladder that embodied-systems reviewers apply before they trust a demo.4---56# IROS Experiments78Use this before submission when the empirical story is not yet locked. At IROS, experiments exist to9prove a robot did something reliably, under stated conditions — not to top a leaderboard.1011## Experiment audit1213- Map each claim to the evidence that supports it: a real-robot trial set, a simulation study, an14 ablation, or a field deployment. A claim with no matched evidence is an overclaim.15- Define the **success criterion before running**, in one sentence a skeptic would accept, and hold to16 it; a criterion invented after seeing results is a reviewer red flag.17- Report **trial counts and resets**: how many attempts, how start conditions were randomized, and what18 reset happened between trials. Hidden resets inflate reliability.19- Publish a **failure taxonomy** with counts, not just a success rate; the failures are where reviewers20 calibrate trust.21- Make **baselines fair**: run them on the same platform, sensors, and route, with comparable tuning22 effort, or state precisely why not.23- Separate **simulation from real** and report the **sim-to-real gap** as a number; an implied zero gap24 is the fastest way to lose a reviewer.2526## The evidence ladder2728| Claim altitude | Evidence IROS expects | Reject pattern avoided |29|---|---|---|30| "The system works" | Trials with n, success interval, and resets stated | "One hero run shown as if typical" |31| "It is reliable" | Failure taxonomy with counts across conditions | "Success rate with no failures reported" |32| "It transfers" | Real-robot numbers plus the measured sim-to-real gap | "Sim results implying real performance" |33| "It beats prior work" | Same-hardware baseline on the same task | "Comparison against a weaker or re-tuned baseline" |34| "It runs onboard" | Measured rate and power under the real compute budget | "Real-time asserted, never measured" |3536## Small-n statistics for real robots3738Real trials are expensive, so n is small — but small n does not excuse a bare mean. Report a success39count as a proportion with a confidence interval (a Wilson interval behaves better than normal40approximation at small n), and for paired system-vs-baseline comparisons on the same trials, prefer a41paired test over independent means. State n every time; "usually succeeds" is not a measurement.4243## Vignette: a manipulation reliability study4445Suppose the system claims reliable grasping of unseen objects. The matching plan: fix an object set and46a success criterion (lifted and held 3 seconds), run a stated number of trials per object with47randomized poses, log every failure by cause (slip, mis-localization, collision), and report the48per-object and pooled success rates with intervals. A simulation sweep over object mass then maps where49the grasp model degrades, and the real-vs-sim gap is stated — every panel tied to a specific claim.5051```text52Trial-logging template (one row per attempt):53 trial_id, object/scenario, start_pose_seed, outcome{success|fail},54 failure_cause, reset_type, wall_time, notes55Aggregate to: success rate + interval per condition, failure histogram, sim-to-real delta.56```5758## Reporting floor5960- Every reliability figure carries n and a criterion; every timing claim carries a measured rate and the61 compute/power budget it ran under.62- Report the compute actually consumed on the robot, not a desktop proxy.6364## Output format6566```text67[Experiment readiness] strong / adequate / weak68[Claim -> evidence map] <claim: trials/sim/ablation/deployment>69[Missing evidence] <trials/resets/failures/baseline/transfer>70[Statistics] criterion set? interval reported? paired where paired?71[Decision-critical next run] <one experiment on the robot>72```
brycewang-stanford