Why this release matters
Collecting high-quality tactile and bimanual real-robot demonstrations at scale is rare. This dataset supplies dense, temporally aligned multimodal traces — vision, proprioception, joint torques, precise per-fingertip force/torque readings, and language decompositions — that make real-world visuomotor and tactile representation learning feasible without sim-to-real gaps from synthetic data.
What Sets It Apart
- High-density real touch: per-frame 60D fingertip force/torque values plus two tactile video streams let models learn both compact and spatial tactile representations, supporting visual‑tactile fusion beyond simple contact flags.
- Bimanual dexterity at scale: two 7‑DoF arms with 22‑DoF hands produce rich coordinated manipulation behaviors across 59 tasks and many long‑horizon episodes, enabling research on coordination, hierarchical policies, and skill composition.
- Language + temporal grounding: every episode has global instructions and 224k temporally grounded subtask segments with 45 skill labels, making language‑conditioned policies and phase‑aware learning straightforward.
- Practical export & tooling: provided in LeRobot v3.0/v2.1 layouts with Parquet frame data and per-camera MP4s to integrate into common ML pipelines.
Who should use it, and tradeoffs
Great fit if you are developing or evaluating imitation learning, vision‑language‑action models, visual‑tactile fusion, hierarchical policy learning, or long‑horizon multi‑phase controllers using real‑robot data. The dataset reduces reliance on simulation and supplies real contact dynamics that are hard to simulate.
Look elsewhere if you need a permissive commercial license (this release is CC BY‑NC 4.0), minimal storage requirements, or a standardized benchmark split — experiments should report season/task splits to avoid temporal leakage. Also, some exported fields (e.g., TCP rotation conventions) are not fully documented and high‑resolution tactile videos increase I/O and storage demands.
