HandX launches a foundation dataset for scalable bimanual motion generation

Strong research release with a clear dataset-first thesis: if you want better hand motion, you need better data, better annotations, and hand-specific metrics.

• –The main value is not just the model benchmark, but the dataset and annotation pipeline that make bimanual motion modeling more tractable at scale.
• –The LLM-based decoupled annotation approach is practical: it turns low-level motion cues into richer semantic supervision without requiring fully manual labeling.
• –The scaling result is the important signal here. It gives a concrete reason to expect better motion coherence as data quality and model size increase.
• –This is especially relevant for robotics, teleoperation, animation, and avatar systems where hand contact timing and inter-hand coordination matter more than whole-body pose alone.
• –The project reads like infrastructure for a research area that has been under-served relative to full-body motion generation.