Questions, answered.

Paddy builds standardized data-collection infrastructure for Physical AI companies training vision-language-action (VLA) models. We make the Harvester, a dual-arm teleoperation rig, and run data collection on it. Standardized hardware, schema and operator protocol mean every dataset is clean enough to train on and consistent enough to compound across engagements.

Physical AI teams that need high-quality manipulation demonstrations without building collection infrastructure themselves. We are currently working with companies in lab automation.

Our team is in New York, NY. We also run a real2sim2real testing lab in San Francisco, CA.

A portable dual-arm teleoperation platform that travels to your lab or facility. It pairs two UFactory xArm 7 arms (7 DoF each, 3.5 kg payload, 700 mm reach, ±0.1 mm repeatability) with an extruded-aluminum frame, leveling casters, adjustable height and 90° or 45° arm mounts. The teleoperation capture stack is built in-house, not adapted from a vendor SDK.

The Harvester starts at $28,000, depending on configuration. That includes two UFactory xArm 7 arms, the Harvester frame and mobile base, and a camera mounting kit for Intel RealSense. Accessories cost extra. Control software is included; compute is not, so you bring your own workstation.

Reserve with a $1,000 deposit through secure Stripe checkout. The deposit is fully refundable and is applied to your final invoice. Reserve a Harvester →

After you reserve, we confirm your configuration with you and then share a delivery timeline.

The stack runs on ROS 2 Humble, on native Ubuntu 22.04 or on Ubuntu 24.04 via Docker. You provide the workstation.

Every run captures multi-view Intel RealSense RGB and aligned depth (one camera per arm plus top-down), joint states at 100 Hz (position, velocity and effort), commanded and achieved end-effector poses, gripper commands and status, and Quest controller input, all hardware-timestamped. See the full topic list on the Technical page.

Each run is saved as its own MCAP rosbag, grouped by session, with HDF5 output as well. Rosbags convert directly into a LeRobot-format dataset that is ready to upload to Hugging Face. For managed collection, we deliver in your preferred format.

Operators use standard Meta Quest controllers. The headset sits on the desk as a tracking reference, so nobody wears it through hours of collection. Switchable scaling profiles move between a slow, precise mode and a fast repositioning mode, and the arms follow in Cartesian space using UFactory online trajectory planning.

Yes. You define the task and environment. We handle task and success-criteria design, a trained operator team, per-episode QA review and delivery of a ready-to-train dataset.

Yes. Rent Harvester rigs by the day or week at our San Francisco lab to collect a pilot dataset, validate a policy on real hardware or evaluate the platform, with no hardware commitment. Operator support is available.

Use the contact form, email contact@paddydata.ai or schedule a call. We reply within one business day.