For contact-rich manipulation, vision often misses the moment that matters: whether a connector has seated, a latch has snapped, or an object is beginning to slip. Recording touch continuously alongside both hands’ camera views makes those events observable in the training data. A key distinction is that the demonstrations come from handheld grippers, so the data is not tied to one robot’s joint space—but it also does not provide robot execution trajectories.
What Sets It Apart
- Four fingertip visuotactile streams accompany synchronized wrist-camera views, making vision–touch fusion and contact-event learning central rather than optional.
- Episodes cover office, home, and workbench tasks, including insertion, fastening, packaging, and deformable-object handling; this breadth exposes policies to varied contact modes.
- The collection is released task by task. The card reports 101 public sample episodes, while much larger processed and planned totals are not yet publicly available, so current availability is far smaller than the headline scale.
- The data uses LeRobot v3.0 and CC BY-SA 4.0. The license permits commercial use but requires attribution and share-alike terms for distributed derivative datasets.
Who It’s For
A good fit for researchers studying visuotactile representation, contact-rich imitation learning, or cross-modal prediction who can work with handheld demonstrations and adapt them to their own robot. Look elsewhere if you need robot-collected joint states, commanded-action trajectories, or a complete large-scale release today: the collection has no robot execution data, and public samples are only a small subset of the stated pipeline and target.
