PalmDex — Embodiment-Agnostic Tactile Manipulation Dataset
PalmDex is a multimodal robotic manipulation dataset featuring synchronized dual-camera video, dual-hand tactile sensing, and hand pose tracking across diverse real-world environments, collected via human teleoperation. Released in June 2026 under CC BY-SA 4.0, it is one of the only large-scale datasets providing dense tactile sensing data alongside RGB video for manipulation tasks. The embodiment-agnostic design — data is collected using human hands rather than a specific robot platform — enables retargeting to any dexterous robot hand. Demonstrations are annotated at the action-segment level with rich categorical labels covering grasp type, contact region, applied force range, and manipulation phase. PalmDex addresses the critical gap in tactile data availability identified by McKinsey as a structural bottleneck in the humanoid robot supply chain.
| Year | 2026 |
|---|---|
| Embodiments | human hands, dexterous robot hands |
| Modalities | rgb, tactile, proprioception |
| Task categories | manipulation, dexterous-manipulation, pick-and-place |
| Data format | mp4, hdf5, json |
| License | CC BY-SA 4.0 |
| Access | open — commercial use permitted |
| Maintainer | PalmDex Research Team |
| Origin country | US |
What is it?
PalmDex is a multimodal robotic manipulation dataset featuring synchronized dual-camera video, dual-hand tactile sensing, and hand pose tracking across diverse real-world environments, collected via human teleoperation. Released in June 2026 under CC BY-SA 4.0, it is one of the only large-scale datasets providing dense tactile sensing data alongside RGB video for manipulation tasks. The embodiment-agnostic design — collecting via human hands rather than a specific robot — enables retargeting to any dexterous robot hand.
Who is it for?
Researchers working on contact-aware manipulation and tactile sensing integration — the capability identified by McKinsey as a critical structural bottleneck in the humanoid robot supply chain. PalmDex provides training data for policies that must reason about contact forces, grasp stability, and surface texture rather than vision alone.
Key specifications
- Modalities: RGB (dual camera), tactile (dual hand), proprioception, hand pose
- Design: Embodiment-agnostic (human hand collection)
- Annotation: Action-segment level with grasp type, contact region, force range, manipulation phase
- Format: MP4, HDF5, JSON
- License: CC BY-SA 4.0 — commercial use permitted
How it compares
PalmDex is the only large-scale dataset combining RGB and dense tactile sensing for manipulation. DexWild (CMU) and DexUMI (Stanford) cover dexterous manipulation without tactile modality. EgoTouch provides tactile data but at smaller scale. PalmDex's human-hand collection makes it retargetable to any dexterous hand platform — including the Inspire Hand targeted by DexUMI.
Limitations and access notes
Human hand data requires retargeting to robot kinematic structure. Share-alike requirement on derivatives (CC BY-SA 4.0). Tactile sensing hardware varies across platforms — transfer to robot tactile sensors requires calibration.
Linked professions
Frequently asked questions
What makes PalmDex unique among manipulation datasets?
PalmDex is one of the only large-scale datasets with dense tactile sensing alongside RGB video. McKinsey identifies tactile sensing as a critical bottleneck in humanoid robot supply chains — datasets with tactile training signal are rare because the sensing hardware is rare. PalmDex's human-hand collection makes it retargetable to any dexterous hand platform.
Can PalmDex be used commercially?
Yes. CC BY-SA 4.0 permits commercial use. Derivative datasets must be shared under the same license.
What is embodiment-agnostic data collection?
PalmDex collects data using human hands rather than a specific robot platform. This produces demonstrations with natural motion and force profiles that can be retargeted to any dexterous robot hand through kinematic mapping — avoiding hardware lock-in to a single robot type.
What tactile information does PalmDex capture?
PalmDex captures continuous pressure maps from wearable tactile sensors across both hands simultaneously, alongside hand pose tracking at joint level. Annotations include grasp type (pinch, power, lateral), contact region (fingertip, palm, lateral), applied force range, and manipulation phase (approach, grasp, manipulate, release).
How do I access PalmDex?
PalmDex is available on Hugging Face at huggingface.co/datasets/PalmDex/PalmDex. No registration required.