MSK-Bench Weights and Experiment Artifacts

The data/ directory contains policy weights for MyoSuite/MSK-Bench tasks, together with their configuration files, training logs, and evaluation results. The top-level folders identify algorithms or experiment series. Weights from different folders are not interchangeable, even when their task names look similar.

Where to find weights

The word best is part of the original filenames; this README does not infer which metric was used to select those checkpoints. To reproduce an evaluation, also check the configuration, environment version, and any associated preprocessing or normalization state.

deprl/: DEPRL baselines

This directory has three task folders. Each numbered subdirectory contains one experiment run. The environment names below come from the corresponding config.yaml files.

A run typically contains:

• config.yaml: environment, DEP, MPO, training, and checkpoint settings.
• script.py: the saved training script, useful when inspecting or reproducing the experiment.
• checkpoints/best.pt: the checkpoint labeled best; step_*.pt files are checkpoints saved at specific training steps.
• checkpoints/logger.pt and checkpoints/time.pt: saved logging or training-progress state, not standalone policy weights.
• log.csv: training and test metrics.
• eval_*.png and *_compliance_report.csv: plots and numerical reports for energy, jitter, muscle actions, and physical joints.
• EMG_*: electromyography-related results, where present.

msgym/: Policies grouped by environment and run

The first level is a MyoSuite environment, such as myoLegWalk-v0/ or MyoRun-v0/. The next level, such as 0625-005737_0/, contains the artifacts from one run. Some environments have several runs; keep a model with the configuration and environment state from the same run. All current run configuration JSON files identify the agent as SAC_DynSyn.

Environment folders:

MyoBunnyHop-v0       MyoCatch-v0            MyoChinUp-v0
MyoCrawl-v0          MyoDoorOpen-v0         MyoHurdle-v0
myoLegBalance-v0     myoLegSquat-v0         myoLegStair-v0
myoLegStandRandom-v0 myoLegWalk-v0          MyoMazeWalk-v0
MyoPoleWalk-v0       myoPowerlift-v0        MyoRun-v0
MyoSidestep-v0       MyoSingleLegStance-v0  MyoSteppingStones-v0
myoWalkAndSit-v0

Common files in a run:

ppo/ and sac/: Task-specific checkpoints and logs

Both directories pair a <algorithm>_<task>_checkpoints/ folder (models and evaluation results) with a <algorithm>_<task>_logs/ folder (evaluation and training logs). For example, ppo_balance_checkpoints/ pairs with ppo_balance_logs/. SAC uses the sac_ prefix.

Current task suffixes in ppo/:

balance, bunny, catch, chinup, crawl, door, hurdle, maze, pole_walk,
powerlift, reach, run, sidestep, sit, slidehard, squat, stair, stance,
stepping_stones, walkandsit

Current task suffixes in sac/:

balance, bunnyhop, catch, chinup, crawl, door, hurdle, maze, pole_walk,
powerlift, reach, run, sidestep, singlelegstance, sit, slidehard, squat,
stair, stepping_stones, walkandsit

These suffixes preserve the original directory names. The two algorithms do not have exactly the same task set or spelling. Key files in each *_checkpoints/ folder are:

• best_model.zip: a Stable-Baselines3 policy archive.
• vec_normalize.pkl: vector-environment normalization state associated with that task's model; check and use the matching file when evaluating.
• eval_*_energy_plot.png and eval_*_jitter_report.png: energy and jitter evaluation plots.
• Muscle_Actions_compliance_report.csv and Physical_Joints_compliance_report.csv: per-dimension muscle-action and joint statistics, including derivative-energy and jerk fields.

The matching *_logs/ folder typically contains eval_logs/evaluations.npz (evaluation data) and events.out.tfevents.* under PPO_1/ or SAC_1/ (TensorBoard training logs).

residualrl/: ResidualRL experiments

The run/, stair/, and walk/ folders correspond to MyoRunResidual-v0, MyoStairResidual-v0, and MyoWalkResidual-v0 in their configurations. Each task folder contains:

• config.yaml: environment, DEP, MPO, and training parameters.
• script.py: the saved training script.
• checkpoints/best.pt: the checkpoint labeled best.
• checkpoints/step_*.pt: checkpoints saved at specific training steps; walk/ retains more intermediate steps.
• checkpoints/logger.pt and checkpoints/time.pt: logging or training-progress state.
• log.csv: training and test metrics.

Keeping files together

1. Choose the algorithm and environment/task first, then select a weight file inside that directory. Similar task names do not make weights from different algorithms interchangeable.
2. For ppo/ and sac/, check best_model.zip together with the task's vec_normalize.pkl. For msgym/, keep the model with the JSON configuration and environment state from the same run.
3. Loading .pt, .pkl, or msgym's *_env.zip files may require the corresponding Python classes and library versions. These are experiment artifacts; the weight files alone are not guaranteed to provide a working inference setup.
4. log.csv, *.npz, *.monitor.csv, events.out.tfevents.*, and the evaluation plots are for inspecting training or comparing experiments. They are not policy weights.