Finevideo VLA Agent: Human 3D-Pose Action Tokens for Robotics

EmpathicRobotics is creating a CC-BY-SA permissive world model dataset for human 3D-pose action tokens, inspired by a 2025 research paper. The dataset is based on the FineVideo dataset and aims to extract human pose, facial, hand, and foot points for model training. The dataset page was last updated on 2026-04-03.

Use Cases

• Train models for human motion retargeting to simulated and physical robots based on the described 3D-pose tokens.
• Develop world models for robotics based on the CC-BY-SA permissive action token dataset.
• Refine human pose estimation, including facial and hand points, for action recognition tasks.
• Benchmark action token generation methods for human motion, as inspired by the cited research.

Strengths

• Dataset creation is inspired by a specific 2025 research paper (arXiv:2506.06072), providing a clear technical foundation.
• Aims to include comprehensive human pose data, including facial, hand, and foot points.
• Explicitly targets retargeting to specific robotic platforms like the H1 unitree.

Limitations

• The dataset is marked as a 'Work In Progress'; completeness and stability are not guaranteed.
• Row count, column definitions, and file formats are unknown, limiting suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

Based on the FineVideo dataset.

Collection Method

Created by cleanup and extraction of human pose data from source videos.

Time Range

null

Freshness

Last updated 2026-04-03 17:42:51; freshness should be verified.

Geography

null