Data and checkpoints¶

The git repository contains only the code plus the small configuration files. Everything else lives on Zenodo and must be unzipped into the data/ directory before running the training, evaluation or plotting scripts.

Included in the repository¶

No download needed for these:

Directory	Contents
`data/env_configs/`	Per-task environment configuration JSONs (`ENV_CONFIG_PATH`).
`data/expert_configs/`	Expert-policy configuration JSONs (`EXPERT_CONFIG_PATH`).

Downloaded from Zenodo¶

Find and download these from the Zenodo record at https://zenodo.org/records/21493316, then unzip them into data/:

Directory	Contents	Needed for
`data/student_policies/`	Trained OBC / Arnold / single- and multi-task student policy checkpoints (`.zip` + `vecnormalize.pkl`) and their TensorBoard training logs.	Evaluation (`src/benchmark.py`), activation collection (`plotting/collect_activations.py`), and the student / transfer learning-curve plots.
`data/expert_policies/`	(Super-)expert policy checkpoints and their TensorBoard logs (`EXPERT_POLICIES_PATH`).	Expert evaluation (`src/benchmark.py --expert`) and `plotting/plot_rl_finetuning_curves.py`.
`data/final_benchmarks/`	Per-method, per-seed benchmark result JSONs, plus `expert_policies/` result JSONs used as the baseline.	The paper's radar and ablation bar plots, and the expert baseline in every performance plot.
`data/final_benchmarks_extra/`	CSI (`csi_*`), bilateral, and the PPO w/o-rew-norm benchmark result JSONs, plus the cached MT-SAC / MT-PPO learning curves in `mt-curves/`.	`plotting/plot_csi_analysis.py`, `plotting/plot_csi_curves.py`, the PPO ablation plot, and `plotting/plot_mt_algos.py`.
`data/kinesis/`	MuJoCo model assets for the `kinesis` locomotion task.	Any run that instantiates the `kinesis` environment.

Resulting layout¶

data/
├── env_configs/              # in repo
├── expert_configs/           # in repo
├── student_policies/         # Zenodo
├── expert_policies/          # Zenodo
├── final_benchmarks/         # Zenodo
├── final_benchmarks_extra/   # Zenodo
│   └── mt-curves/
└── kinesis/                  # Zenodo

Scripts also write into data/ as they run — data/activations/, data/pca_analysis/ and data/figures/ are created on demand.

Weights & Biases¶

plot_mt_algos.py runs offline

plotting/plot_mt_algos.py reads its MT-SAC / MT-PPO learning curves from the cached CSVs in data/final_benchmarks_extra/mt-curves/ (included in the final_benchmarks_extra download), so it runs without wandb access.

Any missing curve is re-fetched from Weights & Biases automatically and re-cached, which requires a logged-in wandb account with access to the runs referenced in the script. External users may not have access to the original wandb runs, and they may eventually be deleted by the Arnold team.