Month 6 · ~8.5 h of talks + final project

The frontier, and your first robot

Your goal this month: learn where robot learning is heading right now, and build one project that proves everything you learned.

These are guest lectures and seminars from the people building the current generation of robot AI, about 8.5 hours in total.

Then you build. Simulation is enough. A real arm is the upgrade.

month 60/9 · 0%

Watch

Robot Learning (CS231N lecture 17) ↗1 lecture

Yunzhu Li, Columbia

If you watched it in Month 3, watch it again. It will look like a different lecture now.

Advancing Robot Intelligence (CS224R lecture 17) ↗1 lecture

Ashish Kumar, AI lead at Tesla Optimus

How RL moves robot intelligence forward, from the person who led AI for Optimus.

Frontiers (CS224R lecture 18) ↗1 lecture

Chelsea Finn

Where robot learning is heading next.

Gen Control, Action Chunking, Moravec's Paradox ↗seminar

Max Simchowitz, CMU

Mathematical evidence for why the "easy" things are hard for robots.

Ψ0: An Open Foundation Model for Humanoid Loco-Manipulation ↗seminar

Yue Wang, USC

A humanoid that walks and grabs at the same time.

Manipulation Data Pyramid ↗seminar

Yang Gao, Tsinghua

How internet video, human teleoperation and robot experience fill each other's gaps.

Embodied Foundation Models ↗seminar

Jonas Frey, Stanford

Why sim-to-real works for walking robots, and why it's still hard for manipulation.

Robot Learning from Human Experience ↗seminar

Danfei Xu, Georgia Tech / NVIDIA

How robots learn from video and data recorded by humans.

AI and Robotics ↗33 min

Chelsea Finn and Andrew Ng

Ends with a section on how to get into robotics and RL.

Full Stanford Robotics Seminar archive ↗

What to focus on

→Action chunking: predicting a short sequence of actions at once instead of one at a time
→Foundation models for robots, and why they need huge and varied datasets
→The data problem: internet video, teleoperation and robot experience, and what each is good for
→Sim-to-real: why it works well for walking and much less well for grasping

Practice task

Colab notebook coming soon

The final project. Simulation is the main track, and the hardware is an upgrade:

LeRobot docs ↗

train an imitation learning policy on a LeRobot simulated task and evaluate it
change the task slightly (move the object, change its color) and measure how well it generalizes
publish a GitHub repo with a README, a 30-second video and one honest paragraph about what failed
include three numbers: success rate, success rate after one change, and how many demos or steps it took

That second number is the one interviewers care about most. Everyone can make a demo work once. Very few people measure how it breaks.

Hardware upgrade (optional)

The deliverable is the same at every tier. Pick what fits your budget.

Simulation only

LeRobot's simulated tasks on your laptop. This is the main track, and it's enough to finish the path.

$122

One SO-101 follower arm

If you 3D-print the parts. You control it with the code you've written since Month 2.

parts list ↗

$230

Leader + follower pair

You move the leader by hand, the follower copies it, and LeRobot records every movement as training data. Record 50 demos of one task, like a block into a cup, train a policy, and let the arm do it alone.

parts list ↗

Milestone: by the end of this month you can

explain what a robot foundation model is and what data it needs
explain action chunking and why it helped imitation learning
train and evaluate an imitation learning policy with LeRobot
present one robot project with numbers, a video and an honest failure analysis
follow new robot learning research on your own through the Stanford seminar series

Get the next checkpoint by email

One topic, one task and one new resource every week.