Skip to content

feat(action): add human hand pose dataset#71

Merged
mli0603 merged 7 commits into
NVIDIA:mainfrom
hans-yang-tps:data/hand-pose-fd
Jul 20, 2026
Merged

feat(action): add human hand pose dataset#71
mli0603 merged 7 commits into
NVIDIA:mainfrom
hans-yang-tps:data/hand-pose-fd

Conversation

@hans-yang-tps

@hans-yang-tps hans-yang-tps commented Jul 1, 2026

Copy link
Copy Markdown
Contributor

Summary

  • Add the human hand-pose LeRobot action dataset implementation.
  • Register the dataset for action data generation.
  • Add hand-pose domain utility support and normalizer stats.

Testing

  • Not run as part of this PR metadata cleanup; this PR was restored to the existing hand-pose dataset scope after splitting the SFT metadata-filter change into a separate PR.

@hans-yang-tps

Copy link
Copy Markdown
Contributor Author

Need Michael(Zhengyu) Huang's help to review it.

@hans-yang-tps
hans-yang-tps marked this pull request as draft July 1, 2026 09:58
Signed-off-by: Hans Yang <hayang@nvidia.com>
@hans-yang-tps hans-yang-tps changed the title feat(action): add Mecka hand pose dataset feat(action): add human hand pose dataset Jul 2, 2026

@lzhang57 lzhang57 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM!

@mli0603

mli0603 commented Jul 2, 2026

Copy link
Copy Markdown
Collaborator

one more comment. Please fix.

@hans-yang-tps

Copy link
Copy Markdown
Contributor Author

Also removed Mecka keywords.

# Conflicts:
#	cosmos_framework/data/generator/action/datasets/human_hand_pose_lerobot_dataset.py
#	cosmos_framework/data/generator/action/datasets/human_hand_pose_lerobot_dataset_test.py
#	cosmos_framework/data/generator/action/normalizer_stats/human_hand_pose_lerobot_stats.json
#	cosmos_framework/data/vfm/action/datasets/__init__.py
@hans-yang-tps
hans-yang-tps marked this pull request as ready for review July 9, 2026 08:45
@hans-yang-tps
hans-yang-tps marked this pull request as draft July 16, 2026 00:20
@hans-yang-tps
hans-yang-tps marked this pull request as ready for review July 16, 2026 00:28
@hans-yang-tps hans-yang-tps changed the title feat(action): add human hand pose dataset Make SFT metadata filters configurable Jul 16, 2026
@hans-yang-tps
hans-yang-tps marked this pull request as draft July 16, 2026 13:57
@hans-yang-tps hans-yang-tps changed the title Make SFT metadata filters configurable feat(action): add human hand pose dataset Jul 16, 2026
@hans-yang-tps
hans-yang-tps marked this pull request as ready for review July 16, 2026 14:14
@hans-yang-tps

Copy link
Copy Markdown
Contributor Author

@lzhang57 @mli0603 could you please help to reapprove and merge this? Thanks.

@mli0603
mli0603 enabled auto-merge (squash) July 17, 2026 01:19
@mli0603
mli0603 merged commit 5048fb1 into NVIDIA:main Jul 20, 2026
9 of 16 checks passed
lfengad pushed a commit to NVIDIA/cosmos that referenced this pull request Jul 22, 2026
## Description

Add a Cosmos3 Nano functional-dexterity hand-pose inference example to
the Action cookbook, covering both Cosmos Framework and vLLM-Omni
execution paths.

Depends on NVIDIA/cosmos-framework#71 for the public human hand-pose
dataset definition.

## Changes

- Add one self-contained validation episode for the task `unscrew nut
with screwdriver` as a LeRobot asset with one Parquet file and one MP4
file.
- Reindex the public-facing episode metadata and remove source user IDs,
original episode IDs, and private filesystem paths.
- Add a hand-pose action-definition section to the Action cookbook
README.
- Add matching hand-pose setup, inference, and preview sections to
`run_fd_with_cosmos_framework.ipynb` and `run_fd_with_vllm.ipynb`.

## Validation

- Selected the episode from the validation split, not the training
split.
- Asset integrity: 1,440 video frames at 30 FPS, 832x480; one matching
Parquet episode.
- Cosmos Framework notebook: 45 cells, 18 code cells, zero errors, and
zero unexecuted non-empty code cells.
- vLLM-Omni notebook: 38 cells, 17 code cells, zero errors; only the two
pre-existing empty placeholder cells were unexecuted.
- Hand-pose output through each path: 17 frames at 15 FPS, 832x480.
- Notebook sources contain no stored outputs or execution counts.

---------

Signed-off-by: Hans Yang <hayang@nvidia.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants