How the RVPU works
A perception unit — functions, not a compiler toolchain.
The RVPU is a Robotics Vision Processing Unit: a curated set of perception functions that robotics engineers wire cameras into. Pre-processing, ISP, and multi-sensor sync run on the processor; AI-aided perception engines run in FPGA logic, fed by zero-copy DMA. You configure a pipeline and point it at your rig — there is no model to compile, no bitstream to author, and no EdgeAI team to hire.
Preliminary demo target
30+ Hz
Preliminary — target
One pipeline, running on the hardware it ships on.
Depth Anything V2 monocular depth + stereo matching on an AMD Kria KV260, real-time (30+ Hz, ~33 ms per frame) within the KV260 power envelope. Preliminary target, pending our own demo.
What we build
Three decisions behind the RVPU.
- Pillar 01
01
A curated perception feature set
Dense stereo and multi-view 3D from your own cameras and geometry, automatic self-calibration against reference frames to compensate for drift, object detection and segmentation with 3D mapping and temporal tracking, and cross-modality work such as thermal–RGB matching with follow-on inference on the matched regions. Configured through the SDK — camera intrinsics, resolution, thresholds — not authored as model code.
- Pillar 02
02
A camera interface that does not lock your rig
A digital backbone — Ethernet today, denser options under evaluation — plus adapter boards with universal connectors for USB, CSI, MIPI, and LVDS. Attach as many cameras as your platform needs, in the layout your platform needs, without waiting on a core redesign tailored to you.
- Pillar 03
03
Robotics-shaped outputs + forensic trace
Outputs land in the form robotics stacks actually consume: camera-frame 3D positions, depth maps, and tracked objects over ROS 2 and GStreamer. An 8,192-entry cycle-accurate hardware trace buffer, hardware root of trust, signed firmware, and per-module attestation make every run verifiable for mission review and ROE compliance.
Configure and deploy
The SDK turns a rig description into a running pipeline.
Describe your cameras and their geometry, pick the perception functions you need, and stream 3D positions, depth maps, and tracks into ROS 2. Bring a custom detector from PyTorch, ONNX, or HuggingFace when a pipeline needs one. No CUDA in the loop.