The through-line
Spacecraft, launch vehicles, cars and robots all answer the same three questions — and increasingly they answer them with a blend of physics-based estimation and learned models.
Where am I?
Inertial navigation integrates motion and drifts; GNSS and star trackers bound that drift; an EKF fuses them into one estimate carrying an honest covariance.
What's around me?
Cameras, LiDAR and radar feed learned detectors, segmenters and occupancy networks that turn raw returns into a structured, metric scene.
What do I do?
Tracks, semantics and state estimates drive control and planning — and, increasingly, vision-language-action policies that map perception straight to behaviour.
How these pages are written
Derive it, then run it
Every claim is either derived from first principles or measured. Where intuition is easier to build by playing than by reading, the page carries a simulation that runs locally in your browser — no server, no install, sliders you can push until the estimator breaks.
Prose and results, not code dumps
The deep-learning and perception pages explain each model's ideas, tensor shapes and measured numbers. The full runnable implementations live in the bundled Jupyter notebooks, linked from each page, so the narrative stays readable.
Long pages are paginated
The big studies are split into numbered parts with a pager at the top and bottom, and the left-hand contents marks which part each heading belongs to. Deep links still work — a link into a hidden part opens that part automatically. There is a show all on one page switch for printing or searching.
Everything is reachable in two hops
The top bar carries the full site map, the bar underneath it lists the current section's pages,
and every page has its own filterable contents. ⌘K or / opens search over
all pages and all headings of the page you are on.
A taste — estimation, live
Here is the single most important idea in inertial navigation: integrated error grows with time, which is exactly why every real system pairs an inertial core with an absolute reference. Turn on a periodic fix and watch the drift collapse.
Full site map
| Section | Pages | What you'll find |
|---|---|---|
| GNC & Estimation | Attitude · IMU · Filtering · EKF · Star tracker · GNSS · Alt-PNT · NASA practices | Derivations plus simulations for the estimation backbone of autonomy |
| System Analysis | ADCS pointing · Rocket navigation · Perception stack · ROS 2 architecture · QoS & real time · Metrics · Model port · FPS & GPUs · Critique | Full studies: requirements → budgets → control/estimator → FDIR, with real data and honest failure modes |
| Hardware | IMUs · Gyroscopes · Star trackers · LiDAR · GNSS receivers · Edge compute | Grades, specs, part tables and selection guides at datasheet level |
| Deep Learning | DINOv2 · GroundingDINO · RF-DETR · SAM · FastSAM · Stable Diffusion · LLMs · VLA · OpenEMMA | Concepts, tensor shapes and measured results, with the notebooks behind every page |
| Reinforcement Learning | On- vs off-policy · A2C · PPO · GRPO · DDPG · PPO & GRPO for LLMs | Four methods derived, implemented from scratch and measured on LunarLander, then mapped to RLHF |
| Nomenclature | Every symbol used on the site, by domain, with the overloaded ones called out | Start here if an equation uses a letter you do not recognise |
| Algorithms | RANSAC · ICP · Camera model · Stereo · PnP · SfM · Bundle adjustment · Gaussian splatting · EKF-SLAM · FastSLAM · GraphSLAM · Hyperparameter tuning · Optimizers · RRT & RRT* | Robust fitting, multi-view geometry, the three flavours of SLAM and sampling-based planning, each derived with its failure modes — plus how to actually train the learned parts |
| Techniques | Gyro bias cancellation · FDI · LinCov · Joint observability · Schuler oscillation · IMU acceptance testing | Six different answers to "what do you do about an error you cannot measure directly?" |