技能 数据科学 机器人研究可复现性指南

机器人研究可复现性指南

v20260724
icra-artifact-evaluation
本指南为机器人学研究人员提供了全面的方法论,指导如何打包和记录复杂的科研产物(包括算法代码、仿真环境、硬件设计、数据集和实验日志)。旨在确保研究具有高度的可复现性,帮助作者遵循“五分钟怀疑论者标准”,实现从投稿到公开的完整生命周期管理。
获取技能
427 次下载
概览

ICRA Artifact Evaluation

ICRA, unlike several software-systems conferences, has had no standing artifact evaluation committee or badge system in recent cycles (re-check the current year's calls before asserting this to authors — tracks appear and disappear). The absence cuts both ways: nobody will certify your artifact, and nobody will catch its problems before readers do. This skill applies an artifact-evaluator's discipline voluntarily, because in robotics the artifact often is the contribution's proof.

The robotics artifact stack

A robotics paper's artifact is rarely just "the code." Inventory all six layers and decide, per layer, released / partially released / withheld-with-reason:

Layer Typical contents Common blocker
Algorithms planners, controllers, learning code none — release
Integration ROS launch files, configs, calibration "works only on our stack"
Simulation worlds, robot models (URDF/SDF), randomization third-party asset licenses
Learned weights trained policies, perception models training data licensing
Hardware design CAD, PCB, BOM for custom parts patent/commercialization plans
Evidence rosbags, trial videos, session sheets raw-log size (TB-scale)

A paper whose novelty is a custom end effector but which releases only Python scripts has released the wrong layer. Match the release to the claim.

The five-minute-skeptic standard

Package for a reviewer who gives you five minutes before forming a judgment:

git clone <anonymized-artifact-url> && cd artifact
docker build -t icra-artifact .          # or: ./setup.sh — one command, pinned deps
docker run icra-artifact make figure3    # regenerate a headline result in sim
docker run icra-artifact make table2     # recompute stats from released logs
cat CLAIMS.md                            # claim → command → expected output map
  • The first runnable thing must not require a robot: a simulation reproduction or a log-replay analysis gives the skeptic a win in minutes.
  • CLAIMS.md maps each paper claim to a command and its expected output, and explicitly lists which claims require physical hardware (audit tier — see icra-reproducibility).
  • Pin everything: base image, ROS distro, Python deps, simulator version. "Latest Gazebo" is a bug report generator.

Hardware-dependent artifacts

For layers that need the physical robot:

  • Ship a log-replay mode: the perception and decision stack runs against released rosbags, letting readers verify the pipeline without the arm.
  • Ship the sim twin: the same launch files targeting the simulated platform, clearly marked as not the source of the paper's hardware numbers.
  • Document the hardware interface narrowly (which driver topics/services the stack expects) so ports to other platforms are feasible.
  • For custom mechanisms, releasing CAD + BOM converts "trust our gripper" into "build our gripper"; if commercialization blocks this, say so in the paper rather than staying silent.

Datasets and logs as artifacts

  • Extracted per-trial features plus a representative raw sample beat an undifferentiated terabyte dump; provide a download script with checksums.
  • License explicitly (CC-BY for data, permissive or copyleft choice for code); robotics data with humans in frame needs consent/ethics notes.
  • Long-term hosting: university archives, Zenodo-style DOI services, or IEEE DataPort outlive lab NAS boxes and personal cloud links.

Review-time vs acceptance-time states

Under the double-anonymous policy (2026 cycle onward), the artifact has two lives:

  1. Review state: anonymized hosting, no org/usernames in URLs, git history squashed (history leaks author emails), license file present but copyright-holder line deferred ("held for anonymity").
  2. Public state (post-acceptance): real hosting, authors restored, DOI minted, README linking the IEEE Xplore entry, camera-ready pointing at the permanent URL (coordinate with icra-camera-ready).

Prepare both from the start; converting a name-riddled repo to anonymous form in deadline week always misses something.

Safety as an artifact property

Robotics artifacts can move mass. Before anyone external runs your controller:

  • Ship conservative default limits (velocity, torque, workspace bounds) and make the paper's aggressive settings an explicit opt-in flag.
  • Document the E-stop assumption and any human-proximity constraints in the hardware protocol, not just the lab's tribal knowledge.
  • State firmware/driver versions known to behave; a controller tuned on one firmware can oscillate on another, and that failure lands on your artifact's reputation.

Withholding honestly

Legitimate reasons to withhold layers exist (industrial partners, export controls, safety of a hazardous procedure). The rule is disclosure: state in the paper what is withheld and why, and maximize the released remainder — e.g., withheld weights but released training code and evaluation harness. Undisclosed gaps discovered later cost more reputation than declared ones.

Packaging sequence

  1. Inventory the six layers; mark release state and blockers for each.
  2. Build the Docker/pinned environment; verify make figure3 on a clean machine.
  3. Write CLAIMS.md; separate rerunnable from hardware-audit claims.
  4. Add log-replay mode and sim twin for hardware-bound layers.
  5. Produce the anonymized review state; leak-check URLs, history, metadata.
  6. Stage the public state for acceptance day.

Output format

[Layer inventory] released: <layers> | partial: <layers> | withheld: <layers+reason>
[Five-minute test] clean-machine run of headline command: pass / fail
[CLAIMS.md] complete: y/n — hardware-only claims flagged: <list>
[Replay/sim twin] present: y/n
[Anonymity state] review-safe: y/n — leaks: <list>
[Hosting] review URL type + acceptance-day plan
信息
Category 数据科学
Name icra-artifact-evaluation
版本 v20260724
大小 6.2KB
更新时间 2026-07-28
语言