Sswa12/thesis-nwm-checkpoints

NWM Checkpoints — CDiT Navigation World Model

Pre-trained checkpoints for the CDiT (Conditional Diffusion Transformer) Navigation World Model. Used for counterfactual trajectory rollout in the safety-critical CAST pipeline.

Checkpoints

File	Model	Params	Training
cdit_xl_ego4d_200000.pth.tar	CDiT-XL/2	~600M	200k steps, Ego4D + navigation (best generalisation)
cdit_xl_100000.pth.tar	CDiT-XL/2	~600M	100k steps, navigation only
cdit_l_100000.pth.tar	CDiT-L/2	~300M	100k steps
cdit_b_100000.pth.tar	CDiT-B/2	~100M	100k steps
cdit_s_100000.pth.tar	CDiT-S/2	~33M	100k steps

Recommended: cdit_xl_ego4d_200000.pth.tar for best visual quality.

Usage

from cast.unsafe_cf.nwm_rollout import NWMRollout

nwm = NWMRollout(
    nwm_repo_path="/path/to/nwm",
    ckp_path="/path/to/cdit_xl_ego4d_200000.pth.tar",
    dataset_name="recon",
)
frames = nwm.rollout(context_images, action_deltas, n_steps=8)
# frames: list of 8 JPEG bytes showing predicted future egocentric views

See: VasuDevi/CAST — cast/unsafe_cf/nwm_rollout.py