IROS Artifact Evaluation
IROS has no formal artifact-evaluation track or badge, so "artifact" here means the evidence bundle a
reviewer could inspect and that a careful author prepares anyway. The audience is an embodied-systems
reviewer who trusts a logged trajectory over a claimed one, and who cannot rerun your robot but can read
your logs, configs, and code.
The robotics artifact stack
-
Hardware ledger: platform, sensors (make/model/firmware), actuators, onboard computer, and power
budget — the embodiment the result depends on.
-
Logs: time-stamped rosbag-equivalent recordings for representative trials, so a claim about a
contact force or trajectory is auditable.
-
Configs: the exact parameters, calibration, and launch files that produced the reported runs.
-
Code: the system as run, with an entry point and a one-minute orientation in the README.
-
Data: any dataset or object set, with provenance and licensing; for restricted data, enough detail
for credible reproduction without violating terms.
What an IROS reviewer inspects first
| Claim type |
First artifact inspected |
Common failure caught |
| "Runs onboard at rate X" |
Timing logs and the compute/power spec |
Rate measured on a desktop, not the robot |
| "Reliable across trials" |
The trial log and failure record |
Success rate with resets and failures omitted |
| "Transfers from sim" |
Paired sim and real logs |
Only sim logs exist; the gap is asserted |
| "Beats prior system" |
Baseline configs on the same platform |
Baseline run with different hardware or tuning |
Because a reviewer will read a log far sooner than rebuild a robot, make the logs and configs
legible first, and polish the code second.
Review-time versus acceptance-time states
-
Review time (double-anonymous): strip organization names, cluster paths, calibration files tagged
with a lab, commit authorship, and any URL carrying a lab identity. If you link code, use an anonymized
mirror or omit the link and describe the artifact.
-
Acceptance time: replace anonymized stand-ins with a public, licensed, citable release; test every
link from a logged-out browser; and add the DOI or archival link the IEEE Xplore record can point to.
Worked vignette: packaging a field-robot result
A submission claims a lidar-inertial system runs onboard across three outdoor sites.
- Ship one representative rosbag per site plus the launch/config that processed it, so a reviewer can see
the sensor streams and the estimated trajectory line up.
- Record the compute-and-power measurement method, not just a headline rate.
- Emit the trajectory-error tables directly from the logged runs so the paper numbers and the artifact
cannot drift apart.
- State which site is easy and which stresses the system, since that mapping is what a reviewer grades.
Artifact layout:
/hardware.md platform, sensors, compute, power
/logs/<site>/ time-stamped bags for representative trials
/config/ calibration, params, launch files
/src/ system as run + README (1-minute orientation)
/data/ provenance + license (or access instructions)
Output format
[Artifact role] review-time anonymous / acceptance-time public
[Contents] <hardware/logs/configs/code/data>
[Anonymity risks] <paths/metadata/URLs/calibration tags>
[Auditability] logged / scripted / described / asserted
[Fixes before upload] <ordered list>
