Physical AI runs on data. We collect it right.
One rig collects the data, trains the model, and validates it —
training data always matches the hardware it deploys on.
even more for generalizable VLA models
"Garbage in, garbage out — even at scale. Inconsistent data collection is why most VLA training runs underperform."
Most Physical AI companies collect training data the same way: improvised rigs, inconsistent schemas, no standardized capture protocol. The operator setup changes between sessions. The camera angles drift. The joint state frequency doesn't match what the deployed system expects.
The result is datasets that introduce noise into training rather than signal. You spend months collecting data and the model still generalizes poorly — not because of the model, because of the data.
We built Paddy to close that gap. Standardized hardware, standardized schema, standardized operator protocol. Every collection engagement produces data that's clean enough to train on and consistent enough to compound across engagements.
One rig, three jobs.
Teleoperated demonstrations, captured to a standardized schema and protocol on every engagement.
Data captured on the same kinematics your model will deploy on, so training data reflects the hardware it runs on.
Replay and evaluate policies on the rig that produced the data, closing the loop from collection to deployment.
Improvised collection rigs create a gap between the data a model trains on and the hardware it deploys on. Using one platform across collection, training, and validation closes that gap.
The collection rig, ready to deploy.
Dual-arm teleoperation platform built for specialized environments. Portable — travels to your lab or facility. Matched to whatever you deploy, so the data your model trains on reflects the hardware it runs on. The teleoperation capture stack is built in-house, not adapted from a third-party vendor SDK.
Includes 2× UFactory xArm 7 arms, the Harvester frame and mobile base, and camera mounting kit for RealSense D405 or D435. Accessories available for an additional fee. The $35,000 configuration adds an RTX 5080 workstation, shipped set up and paired with the rig.
Secure checkout by Stripe · Read the FAQ
Three ways to work with us.
We deploy collection rigs and collect the data — that's it. Purpose-built for Physical AI companies that need high-quality manipulation demonstrations without building the infrastructure themselves.
We deploy a Harvester rig at your facility, matched to your production hardware. Standardized data pipeline from day one — no improvised setups, no schema drift between sessions.
- 2× UFactory xArm 7 · 14 DoF
- RealSense D405 or D435 camera mounting
- Full ROS2 data pipeline
- MCAP + HDF5 output, hardware-timestamped
- Control software included — RTX 5080 workstation optional, or bring your own
We run the collection and send you the data. You define the task and environment — we handle operator protocol, QA, and delivery in your preferred format.
- Task and success criteria design
- Trained operator team
- Per-episode QA review
- Joint trajectories, force/torque, grip state, egocentric video
- Ready-to-train dataset delivery
Paddy operates a real2sim2real testing lab in San Francisco with Harvester rigs available to rent by the day or week. Collect a pilot dataset, validate a policy on real hardware, or evaluate the platform before buying a rig.
- Daily and weekly rates
- Operator support available
- Real2sim2real evaluation loop
- No hardware commitment
Ready to talk about your data?
Tell us what you’re building. Book a call to scope a data project, or send a message to pre-order the Harvester or book lab time.