Data
Paddy provides an embodied data infrastructure that unifies data collection, model training, and validation on a single hardware rig, ensuring training data precisely matches deployment hardware. The Harvester platform features two UFactory xArm 7 arms, Intel RealSense depth cameras, and a ROS2 pipeline that outputs hardware-timestamped MCAP and HDF5 files, eliminating schema drift and improving model generalization. Paddy also operates a real2sim2real lab in San Francisco where rigs are available for daily or weekly rental.
- Robotics
- Artificial Intelligence
- Hardware
Funding
Founders
Founder details are not available yet.
Product
Problem
Improvised collection rigs create a gap between the data a model trains on and the hardware it deploys on, introducing noise into training rather than signal. Teams spend months collecting data and the model still generalizes poorly—not because of the model, but because of the data.
Solution
Paddy provides an embodied data infrastructure that unifies data collection, model training, and validation on a single hardware rig, ensuring that training data precisely matches the deployment hardware. The Harvester platform features two UFactory xArm 7 arms, Intel RealSense depth cameras, and a ROS2 pipeline that outputs hardware-timestamped MCAP and HDF5 files, eliminating schema drift and improving model generalization. Every collection engagement produces data that is clean enough to train on and consistent enough to compound across engagements. Paddy also operates a real2sim2real testing lab in San Francisco with Harvester rigs available to rent by the day or week, enabling teams to collect pilot datasets, validate policies on real hardware, or evaluate the platform before purchasing a rig.
Target Audience
Primary customers are Physical AI companies in lab automation, as well as research groups and robotics companies needing standardized data collection, model training, and validation on matched hardware.
Features
- 2× UFactory xArm 7 arms with 7 degrees of freedom each, 3.5 kg payload, 700 mm reach, and ±0.1 mm repeatability
- Intel RealSense D405 or D435 depth cameras with multi-view mounting for comprehensive visual capture
- Full ROS2 data pipeline with hardware-timestamped MCAP and HDF5 output, eliminating schema drift
- Meta Quest controller-based teleoperation with headset-off operation for extended 2-3 hour sessions
- Cartesian online planning with latest-wins command mailbox, validation of stale stamps, workspace bounds, step size, velocity, and tracking error
- Scaling profiles (0.7× precise and 1.2× fast) toggleable mid-motion with per-arm scale logging
- Safety features including deadman grip trigger, 20 Hz teleop heartbeat monitoring, and USB foot switch e-stop
- Real2sim2real lab in San Francisco with daily and weekly rig rental rates