4. Reproduction, Hugging Face and GitHub Release

4.1 Publishing Layering

Public tutorials need to meet the criteria of readability, reproduction, and security simultaneously. It is recommended to break down the resources into three layers:

In GitHub documentation, commands use variables such as $PROJECT_ROOT, $DATA_ROOT, and $MODEL_ROOT. Absolute paths on the author’s machine should not be used.

4.2 Reproduction of External Repositories

Execute from this directory:

bashexport ATEC_REFERENCE_ROOT=/path/to/atec-reference-repos
bash code/external_repo_audit/clone_references.sh "$ATEC_REFERENCE_ROOT"

Expected result:

text$ATEC_REFERENCE_ROOT/logic-tars-atec
$ATEC_REFERENCE_ROOT/zsn2024-atec2026
$ATEC_REFERENCE_ROOT/yma867-atec2026-ril

This step only verifies that the public repository and fixed commit are accessible. It does not mean that Isaac Lab has been installed, nor does it indicate that the policy can be run directly on the current machine.

4.3 Minimum smoke test

Code Layer

bashpython3 -m py_compile code/taskd_reference_adapter.py

It can only prove that the Python syntax and file structure are correct, but it cannot prove that the robot’s behavior is correct.

Interface Layer

bashpython3 code/taskd_reference_adapter.py

It should print 16D input adapted to 24D output, as well as the Task D stage corresponding to different scores. It does not start Isaac Sim, nor replaces the official scorer.

Simulation Layer

In the environment where Isaac Sim/Isaac Lab has been configured, use an environment for playback first:

bashpython tools/atec/list_envs.py
python tools/atec/play_task.py \
  --task ATEC-TaskD-B2wPiper \
  --headless --num_envs 1 --debug

The specific environment name shall be based on the current version of the official repository. For the first time, only the window, robot, camera, action shape, and log are checked; no long training is performed.

4.4 Uploading Content via Hugging Face

It is recommended to create a public Datawhale dataset repository, for example:

textDatawhale/atec2026-simulation-challenge-reproducibility

Recommended directory:

textREADME.md
manifests/
  external_repos.yaml
  checksums.sha256
source/
  taskd_reference_adapter.py
  external_repo_audit/
references/
  logic-tars/
  zsn2024/
  yma867/
logs/
  redacted/
videos/

The third-party source code snapshot must include the original license and upstream commit. If the repository is large, only the manifest and difference patches can be uploaded, and the upstream fixed commit should be linked in the GitHub tutorial.

4.5 Pre-release Checks

bashrg -n -i \
  'hf_[A-Za-z0-9]{20,}|token|password|secret|api[_-]?key|/[h]ome/|/[d]ata/' \
  15-Challenge竞赛/ATEC2026

In the inspection results:

• Tokens, passwords, and private keys must be set to zero;
• Absolute machine paths should only appear in examples where “replace with your own path” is specified;
• Local absolute paths should be changed to $PROJECT_ROOT, $DATA_ROOT, or /path/to/...;
• Private server names, internal IP addresses, subscription URLs, and account information should be removed from logs;
• Large files should use Git LFS or Hugging Face instead of being directly added to regular Git history;
• Each weight and data shard must have a SHA256 hash;
• Licenses, authors, upstream links, and fixed commits of third-party repositories cannot be omitted.

4.6 How Tutorial Readers Should Understand the Results

In the tutorial, evidence is divided into three categories:

1. Reproducible: Source code, syntax checks, file hashes, fixed command outputs, or local videos are directly supported;
2. Author disclosure: Ranks, training time, or accuracy are listed in the public repository README;
3. Not reproducible: Requires Isaac Sim, GPU, checkpoint, or an official scorer; currently, only the reproduction path can be provided.

This can prevent mixing up statements like “code can be downloaded”, “policy can be loaded”, “robot can move”, and “high task score” into one conclusion.