sislaboratory/owm-rl-baselines

OWM RL baselines

Reinforcement-learning baselines for the owm-envs ISS docking task, trained and published by outofthisworldmodel.

These are baselines, not world models. Each entry is the best checkpoint of a finished run, chosen on that run's own wandb history rather than its final step -- PPO's entropy collapses partway through a long run, so the last checkpoint is often well past the peak.

Checkpoints

name	algo	obs	env	trained steps	promoted step	chosen on
owm-iss-numerical-v1-coop-ppo-vector	ppo	vector	iss-numerical	70,000,000	12,500,000	val_return
owm-iss-numerical-v1-coop-terminal-ppo-vector	ppo	vector	iss-numerical	25,000,000	22,500,000	val_return
owm-iss-numerical-v1-noncoop-ppo-vector	ppo	vector	iss-numerical	70,000,000	7,500,000	val_return
owm-iss-numerical-v1-noncoop-terminal-ppo-vector	ppo	vector	iss-numerical	25,000,000	25,001,984	manual

Each lives under rl/best/<name>/ and ships final_model.zip, its vecnormalize.pkl statistics, the run's config.yaml and env_config.yaml, and a promotion.yaml recording which run and step it came from and what it scored.

The statistics are not optional. These policies were trained on normalized observations, so loading final_model.zip without its vecnormalize.pkl sibling scores a transform mismatch rather than a policy -- which is why both files keep the names the loader looks for.

Usage

just eval hf:sislaboratory/owm-rl-baselines/rl/best/<name>/final_model.zip

# per-port, under every definition of a successful dock
just eval-matrix runs/best/<name>/final_model.zip

Provenance

name	run	wandb
owm-iss-numerical-v1-coop-ppo-vector	ppo_70M_near	run
owm-iss-numerical-v1-coop-terminal-ppo-vector	ppo_25M_near_terminal	run
owm-iss-numerical-v1-noncoop-ppo-vector	ppo_70M_near_noncoop	run
owm-iss-numerical-v1-noncoop-terminal-ppo-vector	ppo_25M_near_noncoop_terminal	run