Video courses on end-to-end autonomous driving and vision‑language‑action robotics. Every box names a real layer. Every arrow names a real tensor. Every footer cites the source repo.
Every claim cited to its source paper and repo. No hype.
Seven courses across the two stacks reshaping the physical world — autonomous driving and robot learning.
For a decade, driving stacks were built like assembly lines — perception handing to prediction handing to planning, each seam a place to lose information. These courses teach the architectures that collapsed the pipeline.
The VLA series is built from the five most-starred open robot-learning repos — one course each, covering every major action-decoding paradigm plus the framework layer that ties them together.
“If you have found other surveys of this field too vague or too marketing-y, this will feel different.”
The word “self-attention” lands after you have watched it happen. Animation does the heavy lifting; equations stay sparse and serve the visuals, not the other way around.
Every box names a real layer class. Every arrow names a real tensor. Every footer cites the source repo, the source paper, or the industry announcement it came from.
The driving track needs no transformer knowledge. The robotics track needs no robotics. If a concept is required, it gets taught on first appearance.
Each course is built from a single source paper and its reference implementation, then rendered at 4K. Lesson counts are exact.
Survey the four winning end-to-end architectures — UniAD, VAD, DeepRoute, Huawei — in one focused tour.
Three chapters: industry survey · submodule deep-dive · architecture deep-dive
Master tracking, mapping, motion prediction, occupancy and planning — the five heads of UniAD's BEV pipeline.
Ships with 6 hands-on Colab activities
How one open 7B model turns a camera image and a sentence into robot actions — and beats a 55B closed model.
The PaliGemma backbone, the action expert, flow matching, and fine-tuning openpi on your own data.
How NVIDIA's model fuses a slow VLM with a fast flow-matching action head for humanoid control.
The RoboticsDiffusionTransformer end to end — diffusion policies, the DiT, a unified action space, deployment.
Record demos, train ACT/Diffusion/SmolVLA policies, and deploy to a real SO-101 arm.
Status reflects production state, not public availability. Join the list below to hear when a course opens for enrolment.
One email per course launch — what it covers, how long it is, and the papers it is built on. Nothing else.
Free. Unsubscribe any time.