FOR ENGINEERS WHO UNDERSTAND POINT CLOUDS BUT ARE STUCK ON DEEP LEARNING...
Master the Apex Predator Techniques for 3D Deep Learning
Used by Perception Engineers at Waymo, Aurora, and Top Robotics Companies
The gates closed on February 9, 2026. Next-opening planned Mid-2026. Make sure to join the waitlist to get notified.
Dear Friend,
If you're trying to break into self-driving cars and autonomous robots, and keep failing interviews, being reminded that you're not there yet, then this page will show you how to close that gap.
Here's why breaking into Perception feels impossible right now:
The information exists. Research papers get published. GitHub repos are public. Conference talks are on YouTube. And many companies showcase algorithms and state-of-the-art performance year after year.
And yet... None of it is actually designed to teach you!
Papers assume you already know the fundamentals.
GitHub repos break every time you try to run it.
Tutorials use plug-and-play solutions that don't really explain how architectures are built...
And many engineers get stuck piecing together fragments from a dozen different sources, never quite sure if you're learning the right things or just wasting time on outdated approaches.
Meanwhile, job descriptions at Waymo, Aurora, Cruise keep demanding the same skills:"Experience with LiDARs and production-grade 3D object detection pipelines."
If you're reading this and thinking "I don't have that on my resume, isn't it time to change that?"...
Then realize in the industry, Autoware (powering hundreds of AV startups) deploys LiDAR CenterPoint. Waymo's robotaxis rely on algorithms like SWFormer. In robotics, drones, autonomous systems... It's all 3D deep learning on LiDAR data.
That's the gap this course fills.
It's not about theory, but pure practice. You'll implement the exact algorithms self-driving car teams use in production, from PointNet to advanced Point-RCNN approaches.
Waymo's robotaxis process 3D LiDAR data with deep learning at the core of their perception stack
Here's exactly what you'll master:
MODULE I: Point Based 3D Deep Learning
In the first module, you'll get an understanding of the point-based approaches for 3D classification, and build your first algorithm from scratch to do 3D Classification and Segmentation.
What you'll learn:
- Why standard convolutions fail catastrophically on point clouds (and the only convolution technique that actually works — many teams use this, but it's almost never taught)
- The 2 competing approaches to 3D deep learning that divide the entire industry (point-based vs. voxel-based - - you need to know both, and when each one fails)
- How to export techniques from self-driving cars to video games and medical imaging (many of these are 3D based and could work perfectly with Deep Learning if you know how)
- The euclidean vs. non-euclidean trap that stumps 90% of engineers (get this wrong and your algorithm will fail on real-world data)
- The 3 libraries every perception engineer uses (and the many 'popular' ones that will waste your time, I'll show you which to skip and why)
- The point cloud visualization mistake that wastes hours of debugging (binary format nightmares) and how to fix it
- A block-by-block study of PointNet, the algorithm that started it all (most engineers only read the paper; you'll actually code it and understand why it's still used in production)
- The loss function trick that separates research code from production systems (adapt this and you can port any classification network to segmentation)
- The 3 invariance rules that separate working code from broken prototypes (ignore these and your network will fail the moment point order changes even though the 3D object is identical)
- The real reason PointNet fails on large outdoor scenes, and how PointNet++ solves it with hierarchical sampling (this is what Autoware's CenterPoint algorithm actually uses)
- How to adapt any 3D classifier for semantic segmentation with one architectural change (and why most voxel-based approaches hit a wall here)
- BONUS: The PyTorch Cookbook (16-page cheat sheet showing you how to use PyTorch in 10 minutes or less)
- The spatial encoding trick for fine-grained shape recognition (how to train networks that distinguish a sedan from an SUV, even when point clouds are sparse)
- Why self-driving cars can classify a car from just seeing its bumper (the partial-view problem that killed early 3D networks, and the elegant solution that made it work)
- PROJECT 🔥: Implement, train, and run PointNet from scratch (production-quality code for classification AND segmentation—ready to add to your GitHub)
Now, when I say this, I mean it:
You will understand every single block:
And even code the entire algorithm from scratch:
FOR CLASSIFICATION:
AND 3D SEGMENTATION:
MODULE II: Voxel-Based 3D Deep Learning
In Module 2, you'll master the voxel-based approach that powers most production perception systems, including how to implement the sparse convolution techniques companies like Apple pioneered.
What's included:
- The 2 voxelization strategies that divide the industry (occupancy vs. density - you need to know when each approach fails)
- Why Occupancy Networks are overhyped, and the one use case where they actually outperform traditional voxel approaches (most engineers get this backwards)
- The one reason people choose to convert point clouds into Voxels, and the biggest drawback of voxelization networks
- How self-driving car teams actually decide between point-based and voxel-based architectures (the memory vs. performance tradeoff that determines production systems)
- The one mistake even LiDAR professionals make when working with voxels (this resolution error causes false negatives in object detection, I'll show you the simple check that prevents it)
- When 3D CNNs actually outperform point-based networks, and the layer depth sweet spot (go deeper and you're just burning compute for no gain)
- The 'almost unknown' sparse computing trick that avoids wasting 90% of your compute on empty voxels (standard convolutions multiply with zeros constantly — this optimization is how production systems stay fast on millions of point clouds)
- The voxel resolution sweet spot used by autonomous driving teams (too sparse and you lose detail, too dense and inference slows to a crawl
- Step-by-step implementation of Sparse Convolutions from scratch (not using a library, you'll code the indexing logic that makes production voxel networks 10x faster)
- Sparse vs. submanifold convolutions: the distinction that breaks most implementations (use the wrong one and your network either loses accuracy or slows to a crawl)
- The subtle difference between sparse and submanifold convolutions, and why both are needed in most networks
- PROJECT 🔥: Build a complete voxelization pipeline from scratch—point cloud to voxel grid to classification (production-quality code for your GitHub)
Something I want to talk about is how we won't just do something classical and boring here, but we'll really think about how experts do.
For example, we won't just look at 3D Convolutions, we will have an entire module on Sparse 3D Convolutions, which are used in Point Clouds, and even Submanifold 3D Convolutions. The understanding of this difference is usually not taught, but if we want to "get" 3D Deep Learning, we'll need to really dive deep into these concepts.
Let's resume:
- Visualize what 3D CNNs actually learn (we'll examine feature maps layer-by-layer so you understand what the network 'sees'
- Build 3D Occupancy Grids using Sparse Voxels (the same algorithm Tesla's perception team uses)
- The 3 skills that separate perception engineers from tutorial followers (network architecture intuition, sparse optimization, and debugging spatial failures)
- Implement the MeanVFE encoding layer from PointPillars (this feature extraction technique is used in every modern 3D object detector, understand it and you can get many perception codebase)
- How to decode a company's perception architecture from their job descriptions (the terminology clues that reveal whether they use voxel, point, or hybrid approaches
- PROJECT 🔥: Build a complete voxel-based classifier end-to-end (point cloud → voxelization → 3D CNN → classification — ready for your portfolio)
What to remember:
In this project, you will not only architect your own voxel-based algorithm, a technique tons of companies use in the field...
While many courses are shallow and simply connect plug-and-play solutions from GitHub (something many engineers can already do)...
This course will take you in the depths of 3D Deep Learning libraries - you will build your own submanifold sparse Convolutions, and work like a true 3D Deep Learning Researcher, making you an elite engineer in the field.
MODULE III: Cutting-Edge 3D Deep Learning
After building solid foundations in point-based and voxel-based approaches, you'll master cutting-edge techniques like Graph Convolutions, Point Transformers, and Edge Convolutions...
the algorithms pushing the state-of-the-art forward.
What you'll learn:
- 🎙️ EXCLUSIVE INTERVIEW: The PointPillars inventor (the most cited 3D object detector) & CEO of a deep RADAR startup reveal how to become a 3D Deep Learning Researcher
- Graph Convolutional Networks: the expert-level technique most engineers don't know (and exactly when to use them instead of standard convolutions; get this wrong and your network can't capture local geometry)
- Decode a Graph Convolutional Network from scratch on point clouds (you'll see the edge aggregation logic that lets networks learn from local neighborhoods)
- Dive into a guided PyTorch implementation of a Graph Convolutional Network on Point Clouds
- Master RANDLANET: one of the pillar algorithms in 3D segmentation (this efficient architecture handles large-scale outdoor scene... understand it and you can read any modern segmentation codebase
- Why Transformers haven't replaced PointNets/VoxelNet in 3D perception (and when that might change)
- The 3 architectural changes required when adapting 2D segmentation to 3D (skip these and your auto-encoder will fail on sparse point clouds)
- The range-extension trick for short-range LiDARs (this multi-resolution technique lets your network process 2x the distance, critical for robotics and drones)
- Reverse-engineer the algorithms Apple, Uber ATG, and Nutonomy built (why each company chose their specific approach, and what their architectures reveal about production constraints)
- Case Study: How Apple built Face ID with 3D Siamese Networks (the one-shot learning technique that works with minimal training data and is applicable to rare object detection in AV)
- Study Case: How Apple Engineers Built Face ID — An exploration of 3D Siamese Networks
- Case Study: How 3D medical imaging works (MRI, CT scan segmentation with deep learning — the techniques transfer directly to LiDAR perception)
Advanced Segmentation:
A Graph Convolution Architecture you'll Learn:
3D Medical Segmentation Case Study:
TESTIMONIALS
WHAT OTHER STUDENTS BEFORE YOU SAID...
"Jeremy really knows his stuff!"
"At first, I wasn't sure where to start in terms of the underlying algorithms of developing a multi-sensor perception system and wasn't sure that an online course would answer all my questions, but Jeremy really knows his stuff!
The video tutorials are well explained and the assignments or challenges are well designed in making sure that I understand every bits of the algorithms.
I'm currently applying these new knowledge to build a multi-sensor perception system for next-generation off-road, heavy vehicles Thank you for the unlimited lifetime course!"
De Jong Yeong, Postgraduate Researcher at Munster Technological University
"I think this is one of the best course from Think Autonomous"
"The differences between the 3D representations are very well illustrated. Especially the Voxel part is truly great pictured.
As a busy engineer I very much appreciate the workshops that just work. I typically spend many hours playing with them to get behind the details. Thanks for this course!"
Ivo Germann, Development R&D Engineer for Vision Systems
"This course is really amazing!"
"This course is really amazing!
What makes it more awesome is that it's one of it's kind. I cannot image learning all this from random sources and aggregating it all together.
You've made it so easy by putting it all in one place and then providing more resources to explore further.
Thank you for all your efforts and can't wait to learn more."
Ankit Prabhu, Practitioner in Perception for Autonomous Mobile Robots
ADVANCED DLC (Value €199)
LiDAR Object Detection 🔓
After the foundation course, you'll be able to classify and segment point clouds, but you won't be able to detect objects.
And every single perception job at Waymo, Aurora, and Zoox requires 3D object detection experience.
Skills like:
- Anchor Boxes
- ROI Pooling & Dimension Regression
- Foreground/Background Prediction
- Bounding Box Proposal & Refinement
- Two-Stage & One-Stage Object Detection Architectures
This DLC is research-level training. It involves implementing cutting-edge papers (PointPillars, PV-RCNN) with complex libraries like MMDetection3D, OpenPCDet, and more...
This isn't plug-and-play. You'll architect complete detection pipelines from scratch, the kind of work that proves to recruiters you're not just a tutorial follower.
Build your own Anchor Box Generation Algorithm
Don't simply run models, architect them
Implement a State Of The Art Object Detection Project
Build interview-winning portfolio projects (complete implementations that prove you can do the job)
Design Object Detection Solutions
Block By Block
Understand every architectural decision, so you can modify and debug production systems (not just copy-paste code)
Learn Advanced Training Techniques
Master transfer learning and module freezing, the training shortcuts researchers use to train 10x faster by borrowing pre-trained blocks)
Experiment with Multiple 3D Object Detection Solutions
Compare point-based, voxel-based, and hybrid detection architectures, understand the tradeoffs that determine what gets deployed
Become a 3D Deep Learning Researcher
Master research-level techniques like ROI pooling, foreground/background segmentation, anchor refinement (warning: mainly for researchers)
Why The Bundle?
Below are two options to take this course: with or without the 3D Object Detection DLC.
Why I recommend it:
- Near 20% price reduction: The DLC is purchasable on its own after you took the first course, but at the price of 199€
- Seamless progression: The DLC builds directly on foundation concepts (trying to learn object detection without the foundations is like building a house without a frame)
- Perception oriented skill set: Classification + Segmentation + Detection = job-ready perception engineer (not geology or other jobs)
- One-time offer: After this launch, the DLC becomes a separate purchase at full price
FOUNDATIONS EDITION
Master the 3D Deep Learning Foundations
- DEEP POINT CLOUDS
[495€ value]
- LIFETIME ACCESS & UPDATES
- INSTANT ACCESS TO MODULES
Enroll For: 495€
* Or 2 payments of 247.50€
DOMINION EDITION
Master the Foundations & 3D Object Detection
- DEEP POINT CLOUDS
[495€ value]
- LIFETIME ACCESS & UPDATES
- INSTANT ACCESS TO MODULES
- 3D DETECTION DLC
[199€ value]
Total value: 694€
Enroll for Just: 575€
*Or 2 payments of 287.50€
Frequently Asked Questions
How do I know if this course is for me?
The last thing I want is you joining this course and then realizing it wasn't for you.
So let me tell you exactly who I think should NOT join this course:
- If you have no knowledge of LiDARs and Point Clouds, have never opened a 3D visualizer, and don't want to bother building these fundamentals before going "cutting-edge 3D deep learning", don't join.
- If you don't have the patience to search on Google and do side quests, don't join (this course will not reexplain everything; I may use a BatchNorm operator or RELU activation and not explain what it is)
- If you don't want to learn how to process Point Clouds with Deep Learning, don't join (feels obvious, but make sure this is something you do want to learn and spend time on)
- If you're a tourist who just want to master self-driving cars in 2 hours, don't join
Sounds good?
Okay, so let's now see who I built the course for:
- If you validate the prerequisites, and feel a need to master more advanced architectures, this course is for you
- If you want to build a rare and superior profile, that only a few possess, this course is for you
- If you want to learn how to use Deep Learning with Point Clouds, Voxels, Meshes, and in general, 3D Data, then this course is for you
- If you want to become a self-driving car or Perception engineer, by using LiDAR as a specialty, then this course is for you
- If you target less common, and more specialized jobs, in industries like Self-Driving Cars, LiDARs, Robotics, 3D, Drones, this course is for you
- If you want to learn all the above listed in just a few hours, rather than spending months browsing for tutorials without a goal, then this course is for you
My goal is that you begin the course now, and get out of it by the end of the week with strong 3D Deep Learning skills. Not next month, not after a 6 month program.
This week.
It may look short, but it's only because I assume you already have prerequisites. Besides, you don't want you to spend 50+ hours on this topic, and neither do I.
You can expect this course to take between 5 and 8 hours. The LiDAR Object Detection DLC is an additional ~3-5 hours, highly recommended.
This is a self-study online course, which contains videos, articles, drawings, paper analysis, code, projects, and more...
Yes, this course (like all my courses) is hosted on our 2.0 platform, optimized for support, answers, and community learning. You might very well ask questions there and get very quick answers from me or other students.
This course is an intermediate/advanced course, which means several prerequisites are needed, including:
- Understanding of Point Clouds and LiDARs (ideally, you already opened, visualized, or even processed a point cloud file)
- Coding in Python
- Knowledge of Deep Learning using Computer Vision (CNNs, Object Detection, Segmentation, Convolutions)
- PyTorch is a Plus, but what's really required is prior experience coding neural networks
If you feel like you don't validate some of these, I would encourage you to first work on them and come back to this course later.
YET:
Please, don't push back learning these skills, burrying the hard task after 50 other things to do.
Do you want to learn the Perception algorithms used in autonomous robots? Are you committed to making it a pillar of your career?
If yes, then I highly recommend you select a plan (and if you can - the dominion edition is insanely good), and rush to building your future career!
© Copyright 2026 Think Autonomous™. All Rights Reserved.