Perception and autonomy — robots, inspection systems, and models that have to run on the machine.

Wafer defect inspection

Industrial perception · HyVision

Detection, segmentation, and anomaly models for 8-inch and 12-inch wafers, deployed on industrial GPUs inside the inspection tool.

In production · high precision · low false-positive rate

Consumer electronics inspection

Industrial perception · HyVision

An end-to-end AOI pipeline — segmentation, detection, classification — at line speed on glass and metal enclosure parts for a tier-1 manufacturer.

Line speed on AOI equipment · fine-grained defects

autoJetsonBot

Mobile robot · ROS 2/Video

A mobile robot and its Gazebo twin, running one ROS 2 stack. On the robot, a Jetson Xavier NX runs ROS 2 Humble in Docker, and an ESP32 drives the motors and streams encoder and IMU data at 50 Hz into an EKF. slam_toolbox builds maps from the real RPLidar; AMCL and Nav2 navigate the twin, which is built from measurements of the chassis. Web dashboard for teleop and live scans.

Xavier NX · ROS 2 Humble · SLAM on the robot, Nav2 in Gazebo

Camera-only docking

Robot learning · autoJetsonBot

Learning to dock the robot from its camera alone, with a C++ safety gate between every command source and the motors. ROS 2 bags become a LeRobot v3 dataset and a ResNet18 behavior-cloning policy, exported to ONNX and scored by an ArUco judge in paired trials against a classical servo. A stricter judge showed the demonstrations never ended on a stop. In the pre-registered rerun, the policy fell behind the servo from unseen start poses (16/30 against 27/30), and demonstrations recorded through the robot’s 62.2° lens fixed a lens failure (0/30 to 23/30).

Simulation only so far · pre-registered · ROS 2 Foxy, Humble and Jazzy

SmolVLA Edge

VLA inference · LeRobot

Where SmolVLA’s time goes on a small GPU, and what lower precisions and TensorRT buy without changing its actions. The model exports as three ONNX graphs driven by a C++17 runtime on ONNX Runtime or TensorRT, checked against LeRobot at every split point with the real LIBERO checkpoint. On a Jetson Xavier NX one action chunk takes 1.0 s in FP32 and 0.33 s in FP16, with FP16 actions inside the range of the BF16 LeRobot itself runs, once TensorRT is held to the graphs’ FP32 layers.

Xavier NX · TensorRT 8.5 · 3.1× faster in FP16

Seamline

Annotation and training tooling

A desktop app for segmentation: label images, train a model, review its proposals least certain first, and retrain on what was accepted. Remote training runs detached over SSH and survives lost connections; large images open through tile pyramids. Built on mySegLib, a PyTorch library with multi-GPU DDP training, chunked inference, and ONNX export.

Open source · Electron and PyTorch · tested on VisA circuit boards

LabelMeV2

Annotation tooling

A fork of labelme for large industrial images: faster load on big TIFF/JPG, 5000% zoom for small defects, and extra polygon editing. Published on PyPI.

Open source · pip install labelmev2

3D human avatars

3D perception · AIVAR

From a single photograph to a textured 3D face and body, using morphable models (FLAME, SMPL), diffusion-based correction, and neural texture-matching. Deployed on AWS.

Driver monitoring on the edge

On-device perception · Jetson

Head pose, facial landmarks, and seatbelt detection on grayscale day-and-night video, compiled to TensorRT and run on NVIDIA Jetson.

Day and night · TensorRT on device

IP image verification

Retrieval · AIVAR

A vector-search system for animation-character and logo protection: embed a query image, find near-duplicates, return a decision. Built for a government-backed digital IP project.