yitongl/robo_video

Robot I2V Zero-Shot Comparison

This report compares NVIDIA Cosmos3-Nano and Wan2.1-I2V-14B-480P on four HIW-500-LeRobot head-camera episodes.

Inputs are restricted to the left monocular crop of the episode first frame and the global task prompt. No subtasks, actions, wrist camera, subtitles, or watermarks are used for model inputs.

Open index.html in this directory to view condition images, raw model outputs, ground truth clips, and comparison videos.

Episodes

• Episode 4: clothes washing from videos/observation.images.head/chunk-000/file-002.mp4 frame 0 at 0.000000s.
• Episode 22: picking trash to rubbish bin from videos/observation.images.head/chunk-000/file-006.mp4 frame 0 at 0.000000s.
• Episode 25: restocking fridge from videos/observation.images.head/chunk-000/file-006.mp4 frame 3063 at 102.100000s.
• Episode 22983: building children table from videos/observation.images.head/chunk-005/file-909.mp4 frame 0 at 0.000000s.

Notes

• Source head videos were verified as 1280x480 stereo; only the left 640x480 crop was used.
• Episode 25 was extracted from file-006.mp4 at frame 3063, i.e. 102.100000s at 30 FPS.
• Raw model MP4 files are preserved separately from comparison.mp4.
• Cosmos3-Nano ran through Diffusers BF16 with device_map=balanced, no quantization, no offload.
• Wan2.1-I2V-14B-480P ran through the official Wan generate.py I2V path with BF16, official CPU model offload, and no quantization.
• Wan 8-GPU FSDP/USP was attempted first but stalled after DiT shard loading; the completed run uses the official single-GPU offload path.
• FlashAttention could not be installed because the environment lacked nvcc/CUDA_HOME; Wan attention uses a PyTorch scaled dot product attention fallback patched into the local official repo.