Scaling robot training data.

Partnership.

Direct partnership
You define tasks, objects and acceptance criteria. We set up stations, train operators and collect.
Full data rights
Each dataset is collected for one partner and transferred with full rights. We never resell or reuse it.

Operations

Bimanual end-effector states with access to industrial environment.

Delivering scale through two locations working in parallel: data collection and processing in Ahmedabad, India, and robotics software development and verification in California, USA.

Today

18 operators
up to1,000 hours a week

Dataset

Download 50 sample episodes for free.

Bimanual demonstrations recorded by Newlithic operators, published in LeRobot format on Hugging Face.

Get the samples on Hugging Face ↗

Access is gated. Sign in to Hugging Face and accept the dataset terms to download.

Newlithic/bimanual-ee-pose

  • Episodes50 bimanual demonstrations
  • TaskUnscrew an M5 nut from a bolt held by the other hand
  • Per handEnd-effector position in metres, rotation as the first two columns of the rotation matrix (6D), and gripper opening
  • VideoWrist camera view
  • Rate30 fps, video and state synced
  • FormatLeRobot dataset (parquet + MP4)
  • LicenseApache-2.0

Hardware

We can design, source or extend to any end effector or tool.

We can scale the end-effector joint space to a partner's requirement, or use the partner's hardware.

The Newlithic parallel gripper standing on a work surface: yellow 3D-printed jaws with fiducial tags, a black body, and an Insta360 X3 camera mounted on top.
An operator's hand holding the Newlithic gripper. Two yellow parallel jaws carry fiducial tags; a third tag sits on the black body under an Insta360 X3 camera.

Physical constraints

Operators stay inside the robot's reach.

A Newlithic collection station: a white table with a black mat, a pair of yellow grippers, a pile of small bolts, and a clipboard with a printed task card on each side of the table.
Task card · L
Task card · R
Gripper · left hand
Gripper · right hand
Target objects

Pipeline.

Everything inside the camera frame is maximized for our data annotation needs, continuously reducing post-processing at scale.

The Newlithic segmentation and QA tool. On the left, the left glove camera video and right glove mic with a timeline of detected cues. On the right, each spoken start or stop/end cue with its timestamp, the vocabularies that matched it, and buttons to mark it Start, Stop / End or Not a cue.
  1. 01 Wear and record video + rig sensor streams
  2. 02 Perform the task raw session
  3. 03 Segment and label labeled episodes
  4. 04 Track pose + end-effector state
  5. 05 Retarget target robot joint states
  6. 06 Verify verified dataset