技能 人工智能 机器人学习研究可复现性指南

机器人学习研究可复现性指南

v20260724
corl-reproducibility
本指南提供了一个全面的框架,用于确保机器人学习研究的科学可复现性。它详细指导用户如何记录“软件可重现部分”(如模拟器版本、训练配置、代码脚本)和“硬件可审计部分”(如传感器、固件、机械臂规格),确保研究结果可以被专家验证和重建。
获取技能
234 次下载
概览

CoRL Reproducibility

Robot-learning papers have a split reproducibility problem: the training half is software and can in principle be rerun anywhere, while the hardware half is a physical setup nobody else owns. Strong CoRL papers treat these halves differently — the software half is made rerunnable, the hardware half is made auditable — and say plainly which is which.

The rerunnable half: simulation and training

The silent reproducibility killers in this field are version-shaped:

  • Simulator versions change physics. Contact solvers, default damping, and collision margins shift between releases of MuJoCo, Isaac, PyBullet, and friends; a policy's success rate is a function of the simulator build. Pin the exact version and any physics-relevant flags.
  • Environment wrappers drift. Task-suite repositories (benchmark forks, custom reward shims) move under you; record the commit hash of every env repo, not just your own.
  • GPU nondeterminism. cuDNN autotuning and atomics make bit-identical training runs unrealistic — so define reproducibility at the distribution level: same configs + fresh seeds should land inside your reported seed spread.
# repro-manifest.yaml — ship at the repo root; one block per results table
table_3:
  code_commit: ""            # your repo @ exact hash
  env_suite:  {repo: "", commit: "", sim: "mujoco==X.Y.Z", physics_flags: []}
  training:   {config: "configs/table3.yaml", seeds: [0,1,2,3,4], gpu: "1xA100-80GB", hours: 14}
  data:       {demos: "dataset-v2 (1,204 episodes)", url_or_status: "", license: ""}
  checkpoints: {released: true, path: "ckpts/table3/", selection_rule: "last epoch, no eval peeking"}
  evaluation: {script: "eval/run.py", episodes_per_task: 50, init_state_list: "eval/states.json"}
expected_tolerance: "per-task success within ±4 pts of Table 3 mean"

The selection_rule line matters more than it looks: checkpoint selection via test-set peeking is the field's quiet irreproducibility engine, and stating the rule is the cheapest credibility purchase available.

The auditable half: hardware

Nobody will re-run your robot, so the goal is that a skeptical expert could verify the experiment happened as described and rebuild an equivalent rig:

What to document Why an auditor needs it
Robot model, end-effector, firmware/driver versions Behavior differs across firmware, not just robots
Sensor models, mounting poses, calibration procedure Camera extrinsics silently dominate visuomotor results
Control interface: frequency, action space, safety filters "30 Hz end-effector deltas" vs "torque control" are different papers
Deployed compute + inference latency Policy behavior is latency-dependent
Scene inventory: objects (make/size), fixtures, lighting Enables an equivalent-rig rebuild and honest comparison
Raw episode logs and unedited evaluation video The audit trail for every printed success rate

Log every evaluation episode at capture time (timestamped video plus a CSV of outcomes); retrofitting an audit trail after reviews ask for it is impossible.

Data and checkpoint release

  • Demonstration datasets are results: release episode counts, collection method (teleop rig, scripted, crowdsourced), operator count, and filtering rules.
  • Released checkpoints let others reproduce evaluation even when training is too expensive to repeat — for large policies this is often the highest-value artifact you can ship.
  • If data or weights cannot be released (proprietary platform, human-subject footage), say so in the paper with the reason, and release what remains: configs, eval scripts, sim environments, and metrics logs. A precise partial-release statement outperforms a vague "code available upon request."

During review vs after acceptance

  • During review everything must be anonymous (see corl-submission): anonymized repo mirrors, no lab-identifying video, no cloud buckets with named projects.
  • The 2026 camera-ready pipeline has a sharp constraint: PMLR does not accept video as supplementary material, so post-acceptance videos, code, and data live on external hosting (project site, GitHub, archive), linked from the main text (corl.org author instructions, read 2026-07-08). Plan the public homes of artifacts before the October camera-ready deadline, and prefer DOI-stamped archives for anything you cite as permanent.
  • No formal reproducibility checklist was verified for the 2026 cycle in this pack (待核实) — but reviewer expectations enforce one informally; use the manifest above regardless of what the form requires.

Availability statement patterns

Strong:  "Code, training configs, evaluation scripts, and the 1,204-episode
          teleop dataset: <URL>. Checkpoints for Tables 2-4: <URL>. Hardware
          evaluations are documented in Appendix C (rig spec, logs, uncut video);
          the platform itself cannot be redistributed."
Weak:    "Code will be released upon acceptance."      (unverifiable promise)
Broken:  "Results reproducible with standard settings." (no artifact at all)

Pre-submission audit

[ ] repro-manifest present; one block per headline table
[ ] Simulator/env/driver versions + commits pinned everywhere
[ ] Seed policy stated; checkpoint selection rule stated
[ ] Hardware rig spec complete enough for an equivalent rebuild
[ ] Episode-level logs + uncut eval video archived internally
[ ] Data/checkpoint release plan with named blockers, if any
[ ] Anonymous during review; external hosting plan ready for camera-ready

Verify the live cycle's supplementary and camera-ready rules at https://www.corl.org/contributions/instruction-for-authors before promising any artifact channel — PMLR-side constraints and CoRL-side forms both change yearly.

信息
Category 人工智能
Name corl-reproducibility
版本 v20260724
大小 6.12KB
更新时间 2026-07-28
语言