Perception & Computer Vision

The pinhole model, intrinsics, extrinsics, and distortion — and the 30-minute OpenCV workflow that turns a raw USB camera into a calibrated robotics sensor.

~15 min

Before deep learning, computer vision rested on three primitives — corner detectors, descriptors, and matchers. They still power half of modern SLAM and visual odometry. Here's why and how.

~14 min

How pixels move when the camera moves — and how to recover 3D from a moving monocular camera. The geometry that powers visual odometry, SLAM, and every drone that reconstructs a map from a single camera.

~15 min

Three ways to get depth: stereo triangulation, structured light, time-of-flight. What each measures, what each fails on, and when to pick which for your robot.

~12 min

The sensor that took over autonomous driving. Mechanical, solid-state, and FMCW; how to filter raw clouds; ground segmentation, the first step before any object detection.

~13 min

The 2D detection model ecosystem, picked with a robotics lens — latency, small-object accuracy, edge deployment. The runtime constraints that separate paper benchmarks from production.

~13 min

Pixel-precise masks instead of bounding boxes. SAM, Mask2Former, YOLOv8-Seg — what each gives you, and how to pick the right tool for grasping, mapping, or scene understanding.

~13 min

Combining a camera with an IMU for pose estimation that works where GPS fails. Why VIO is its own subfield, and the production-grade systems that fly drones indoors.

~13 min

Latency budgets, domain shift, and the ways robotics CV diverges from ImageNet practice. The reasons your favorite paper's model breaks on a Jetson, and what to do about it.

~12 min

CLIP, GroundingDINO, SAM, and how modern pipelines glue them together for open-vocabulary robot perception. Detect 'the red mug' with no fine-tuning, segment it, grasp it.

~13 min