Senku_WM

Checkpoints for Senku, a self-supervised world model for planning from pixels: each frame becomes a global latent and 49 state tokens. Code, usage and results: github.com/gzee3184/senku_wm.

File	Model
cube/senku_expert_100k.pt	OGBench Cube, LeWM's expert data
cube/senku_mix_100k.pt	OGBench Cube, Cube mix (expert + generated grasp data)
pusht/senku_pusht_400k.pt	PushT (checkpoint chosen on validation)
pusht/senku_pusht_140k.pt, pusht/senku_pusht_250k.pt	PushT, earlier checkpoints (140k = 10 LeWM epochs)
antmaze/senku_antmaze_joints_300k.pt	Visual AntMaze, image + joint readings
antmaze/senku_antmaze_image_200k.pt	Visual AntMaze, image only
cube/senku_mix_40k.pt, cube/senku_mix_idm_40k.pt, cube/senku_mix_idm_100k.pt	Cube mix with and without the inverse-dynamics loss
cube/senku_mix_heads/	optional long-horizon modules for senku_mix_100k
ablations_40k/	leave-one-out ablations, Cube expert data, 40k steps

Each file holds the weights, the model config and the training arguments (no optimizer state).

hf download abrar008/Senku_WM --local-dir ckpts
python eval.py --ckpt ckpts/cube/senku_mix_100k.pt --env cube --seed 42 --holdout-mod 10

License: MIT. Citation: see CITATION.cff in the code repository.