CaPtain-artifacts

This repository holds the small artifacts of CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents. Besides the code and the two models, they are what you need to reproduce the RL run and the paper's evaluations:

• RL start states;
• the warm start of the simulator critic;
• the exact training configurations;
• frozen evaluation inputs;
• three example episodes of CaPtain-4B.

In total there are 1,002 files (about 24 MB). The SFT data are a separate dataset, YijiaFan/CaPtain-SFT-Data.

Pipeline code (SFT, RL)	https://github.com/waltstephen/CaPtain
Harness	https://github.com/waltstephen/CaPtain-Harness
Models	YijiaFan/CaPtain-4B, YijiaFan/CaPtain-4B-SFT

Contents

Path	Files	What it is
critic/warm_start/	5	Warm start of the simulator critic. See Simulator critic below.
starts/libero_fresh_init_rl/{libero_spatial,libero_object,libero_goal,libero_10}/	80	RL start states of CaPtain-4B, for 40 LIBERO tasks. Each task has a <task>.fresh_init state file and a <task>.fresh_init.json sidecar. See Start states.
configs/rl/	7	The RL configuration. See RL configuration. These are the same files as rl/configs/ in the CaPtain repository.
configs/sft/captain-4b-sft.yaml	1	The full SFT configuration of CaPtain-4B-SFT.
eval/tasks/	4	The task lists of the paper's evaluations. See Evaluation task lists. They are the same files as eval/tasks/ in CaPtain-Harness.
eval/libero_pro_postask/	801	Frozen, validated LIBERO-PRO Pos+Task initial states. See LIBERO-PRO Pos+Task states.
eval/libero_max_cases/	80	The 80 LIBERO-MAX case files used in the paper (<suite>-t<id>-<perturbation>-d<level>-p195.json), copied from the frozen LIBERO-MAX release.
examples/captain-4b_libero_official/	23	Three scrubbed CaPtain-4B evaluation episodes on LIBERO official held-out states. See Example episodes.

Simulator critic (critic/warm_start/)

• Model. An ensemble of 5 MLPs (3 × 256). The inputs are 36 physics features, the request index and a task one-hot. It was fitted to final success on SFT-policy episodes from fresh initial states (test AUC 0.907).
• Files:
  • critic_default.pt: the checkpoint the RL configuration loads (critic_dir);
  • critic_split.pt: the train/val/test-split variant used for the reported metrics;
  • feature_spec.json: the feature order, standardisation and advantage rule;
  • metrics_sameT.json: the test metrics;
  • SHA256SUMS.

Start states (starts/libero_fresh_init_rl/)

• Each task has 184 states:
  • 140 wide-layout states, each with its own lighting, camera and material variation (visual in the sidecars);
  • 44 official non-evaluation states (indices 0–10 and 17–49).
• Install them as data/libero_fresh_init of the harness. rl/scripts/prepare_starts.sh then rebuilds the start bank with the reference sha256 df69fab6… (7,360 states); see rl/README.md.

RL configuration (configs/rl/)

• captain-4b.json is the 16-GPU configuration of CaPtain-4B; captain-4b.8gpu.json is an 8-GPU layout.
• staged-reward-spec.json: the staged-reward spec.
• act-ref.json: the action references.
• start-exclude-thin-basket.json: the start exclusions.
• start-pool.reference.json: the reference start-pool manifest.
• official-heldout-states.json: the official held-out state indices. RL never uses these states.

Evaluation task lists (eval/tasks/)

File	Episodes
libero_official_240.tsv	4 suites × 10 tasks × held-out states 11–16 = 240
libero_pro_postask_800.tsv	LIBERO-PRO Pos+Task, 8 cells × 100
libero_plus700_pro160.tsv	LIBERO-Plus 700, plus LIBERO-PRO 4-category 160
libero_max_160.tsv	LIBERO-MAX, 80 cases × 2 arms (Dynamic, Base)

LIBERO-PRO Pos+Task states (eval/libero_pro_postask/)

• 800 .npy files, <suite>_<swap|task>/<task_id>_s<seed>.npy: 8 cells × 100 states.
• manifest.json gives the sha256 of every state and the validity rule (penetration, drift, tilt and drop thresholds). Five invalid starts were replaced by states 17–19.
• Use them with --artifacts <this repo>/eval in the harness's eval/run_jobs.py.

Example episodes (examples/captain-4b_libero_official/)

Episode	Outcome
libero_spatial_0_seed-14	success
libero_10_7_seed-12	long-horizon success
libero_10_8_seed-12	failure at the request cap

Each episode directory holds:

• driver/transcript.jsonl: the model's requests and replies;
• calls.jsonl: every robot call with its receipt;
• capability_manifest.json;
• finish_attempts.jsonl, and completion_gate.jsonl when the gate was used;
• episode_result.json;
• the agent's workspace/TASK.md and workspace/NOTES.md.

Camera images were not retained, and local paths are replaced by placeholders.

Usage

hf download YijiaFan/CaPtain-artifacts --repo-type dataset --local-dir captain-artifacts
# RL start states for the harness:
cp -r captain-artifacts/starts/libero_fresh_init_rl <CaPtain-Harness>/data/libero_fresh_init
# LIBERO-PRO Pos+Task evaluation with the frozen states:
python3 eval/run_jobs.py eval/tasks/libero_pro_postask_800.tsv --artifacts captain-artifacts/eval --out runs/pro --parallel 8

The CaPtain repository's rl/README.md and the harness README.md describe each step.

License

• The CaPtain artifacts (critic weights, configurations, task lists, example episodes) are released under the MIT License, the license of the CaPtain code.
• The LIBERO start states derive from LIBERO (MIT). The LIBERO-PRO states derive from LIBERO-PRO (MIT).
• The LIBERO-MAX case files are copied from LIBERO-MAX (released under CC BY 4.0); please cite LIBERO-MAX when using eval/libero_max_cases/.

Citation

@article{fan2026captain,
  title   = {CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents},
  author  = {Fan, Yijia and He, Xilin and Jiang, Yilei and Wu, Dongming and Wang, Jiaxu and Yue, Xiangyu},
  journal = {arXiv preprint},
  year    = {2026}
}