syCen/droid-action-wm

DROID action-conditioned world model (full fine-tune, step 64730)

This is a Wan2.2-TI2V-5B model fully fine-tuned on DROID 1.0.1 exterior cameras.

• Input: 1 exterior image and 16 end-effector delta actions (5 Hz).
• Output: 16 future frames at 640×480.

file
droid_wm_fullft_s64730.safetensors	all DiT + action-encoder weights (bf16)
action_stats_full_interp.json	action normalisation, required
examples/clip{0,1,2}	held-out-scene clips with reference PSNR for a sanity check

Code, action conventions and the policy-evaluation recipe are in the GitHub repo (see its README).

Held-out-scene accuracy over the 16 predicted frames:

PSNR	SSIM (7×7, RGB)	LPIPS
22.61	0.859	0.082