Solutions · Robotics

Build robot imitation-learning datasets — with the hand–object contact robots actually learn from.

Most 2D annotation tools label objects in a box and stop there. PI Recorder captures the manipulation signal: the hand's pose, the object it is holding, and the grasp or touch contact between them — the data a policy needs to learn a task from demonstration.

Start freeSee pricing

How it works

1 · Record the demonstration

Record video of a hand performing the task, or import existing footage. PI Recorder extracts clean, hand-visible frames automatically.

2 · Detect hand + manipulated object

In Hand + Object mode, MediaPipe finds the 21 hand keypoints and YOLO11 finds the object being held. PI Recorder keeps only the object the hand actually controls — not every object in view — and auto-labels it.

3 · Capture the contact + pose

It computes the grasp/touch contact between the hand and the object, estimates the hand's 6DoF pose and per-finger joint angles, and can recover the object's metric 6DoF via marked corners (PnP) or a device depth sensor (ToF/LiDAR).

4 · Export for training

Export the labeled dataset to JSONL (carrying pose + joint angles + contacts) or to COCO, YOLO and Pascal VOC — the whole dataset or any subset.

What makes the dataset useful for robots

  • Hand pose (6DoF). Estimated position and orientation as both Euler angles and a quaternion, from the 21 hand landmarks.
  • Per-finger joint angles. Kinematic flexion angles for each finger — the hand configuration, not just its box.
  • Hand→object contact. An explicit grasp/touch relation linking the hand to the specific object it manipulates.
  • Metric object 6DoF. Real millimetre distance and rotation via marked corners (PnP) or a phone depth sensor (ToF/LiDAR), not just a scale-ambiguous 2D box.
  • Standard exports. JSONL carries the robotics extras; COCO, YOLO and Pascal VOC are there for the box-only pipeline.

Honest scope

Hand 6DoF from a single camera is a monocular estimate, not calibrated sensor output; object metric pose needs either a known face size (PnP) or a device depth sensor. PI Recorder is explicit about which signal is measured versus estimated — so you can trust what you train on.