Data and checkpoints¶

The git repository contains code and environment/expert configuration files. Benchmark results, model weights, training logs and analysis data are external inputs. The release record is Zenodo.

Environment and expert configurations are already included under data/env_configs/ and data/expert_configs/ without need to download from Zenodo. External inputs from Zenodo use this layout:

External inputs¶

Directory	Contents	Needed for
`data/final_benchmarks/`	Per-method/per-seed result JSONs and expert references; `arnold_single_task/` and `transfer_learning/` logs; `example_training_curve/` base OBC logs; `example_checkpoint/` with one checkpoint and matching normalization file.	Performance, ablation, student, transfer and fine-tuning plots; checkpoint-loading examples.
`data/final_benchmarks_extra/`	CSI evaluation results and `csi*/training/` logs; bilateral and normalization ablations; `mt-curves/` caches; `rl_finetuning/` logs.	CSI, baseline and RL fine-tuning plots.
`data/analysis/`	Curve caches, `reproduction/` metadata, compact hand signals, human/simulation EMG inputs, gait recordings and exported tables.	Learning curves, PCA/NMF summaries, hand and EMG analyses.
`data/final_checkpoints/`	Separately released model checkpoints and associated normalization/configuration files.	Evaluation, fresh recordings and resumed training beyond the bundled example.
`data/expert_policies/`	Expert policy checkpoints (`EXPERT_POLICIES_PATH`).	Expert evaluation and rollout collection.
`data/kinesis/`	Locomotion model assets.	Instantiating the Kinesis environment.

Extract the packages from the repository root:

mkdir -p data
tar -xzf final-benchmarks.tar.gz -C data
tar -xzf final-benchmarks-extra.tar.gz -C data
tar -xzf expert-policies.tar.gz -C data
tar -xzf analysis.tar.gz -C data

The single-task and transfer folders contain training logs and configurations, not model checkpoints. The bundled example checkpoint is data/final_benchmarks/example_checkpoint/rl_model_64670238_steps.zip, accompanied by rl_model_vecnormalize_64670238_steps.pkl. Other checkpoint paths require the separate model release. Training logs alone cannot generate new rollout recordings.

Resulting layout¶

data/
├── env_configs/                    # in Git
├── expert_configs/                 # in Git
├── final_benchmarks/
│   ├── arnold_single_task/          # per-task training logs
│   ├── transfer_learning/           # eight transfer/scratch runs
│   ├── example_training_curve/      # base OBC logs
│   ├── example_checkpoint/          # one model + normalization/configs
│   ├── expert_policies/             # reference JSONs, not expert weights
│   └── <method>/seed_<n>/           # benchmark results
├── final_benchmarks_extra/
│   ├── csi*/training/               # CSI logs beside evaluation data
│   ├── rl_finetuning/               # Arnold fine-tuning logs
│   └── mt-curves/                   # cached MLP baseline curves
├── analysis/
│   ├── reproduction/               # selections and references
│   ├── signals/                    # supplied compact hand recordings
│   ├── emg/                        # human/simulation profiles and gait rollouts
│   ├── historical_curves.csv
│   └── learning_curves.csv.gz
├── final_checkpoints/              # separate download
├── expert_policies/                # separate download: weights
└── kinesis/                        # separate download: model assets

Recordings, fresh intervention results and figures are generated under data/activations/, data/pca_analysis/ and data/figures/, respectively.

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.

Analysis inputs¶

data/analysis/reproduction/ contains selected benchmark filenames, analysis cohorts, EMG muscle mappings and human reference values. Shared task labels live in src/analysis/metadata.py. Recording axes come from file metadata or the task model. Expert summaries under data/final_benchmarks/expert_policies/ provide the reference for relative performance. Offline hand analyses default to supplied compact recordings in data/analysis/signals/. collect_activations.py writes to data/activations/; intervention results use data/pca_analysis/. data/analysis/ holds EMG inputs, exported tables and portable learning curves. The packaged data/analysis/learning_curves.csv.gz and data/analysis/historical_curves.csv files contain cached plot data, not raw recordings.

Collect/export hand recordings with Analysis signals. Human EMG requires external preprocessed subject profiles; EMG analysis documents their format and import command.