Datasets:

---

HuggingFaceVLA

/

community_dataset_v1

like 8

Follow

Hugging Face Vision Language Action Models Research 108

Thanks for publishing this corpus and the PROCESSING_REPORT.md — the transparency made the analysis below
possible.
I've been studying how the camera viewpoint normalization described in §3.2 of the SmolVLA paper ("Camera
viewpoint normalization") landed in the released data, and I have two questions.

1. OBS_IMAGE_3 / slot 3 is populated in ~4% of datasets
   Counting the videos/ feature keys across the 127 datasets in the Feature Remapping Details table:

The paper presents a three-slot scheme (top / wrist / side), but the corpus is overwhelmingly two-camera.
Since smolvla_base declares three visual inputs, could you clarify:

• Is camera3 expected to be provided at inference?
• Given slot 3 was populated in ~4% of pretraining data, is feeding a third view at inference in-distribution, or better left unset?

2. Per §3.2, slot 2 corresponds to the wrist-mounted view (OBS_IMAGE_2). Sampling frame 0 of observation.image2
   across several datasets, the viewpoints look very different from one another:

If slot 2 is meant to be the wrist view, I'd expect these to share a rough geometry, and they don't seem to.
Image2 →

1. https://huggingface.co/datasets/HuggingFaceVLA/community_dataset_v1/blob/main/356c/so100_nut_sort_1/videos/chunk-000/observation.image2/episode_000000.mp4
2. https://huggingface.co/datasets/HuggingFaceVLA/community_dataset_v1/blob/main/Bartm3/dice2/videos/chunk-000/observation.image2/episode_000000.mp4
3. https://huggingface.co/datasets/HuggingFaceVLA/community_dataset_v1/blob/main/CSCSXX/pick_place_cube_1.18/videos/chunk-000/observation.image2/episode_000000.mp4
4. https://huggingface.co/datasets/HuggingFaceVLA/community_dataset_v1/blob/main/Beegbrain/pick_place_green_block/videos/chunk-000/observation.image2/episode_000000.mp4