Physical AI Lab: a physical AI course where every number can be checked

Physical AI Lab is a training series on physical AI (embodied AI) that progresses from introductory to advanced. Each episode consists of a video, an interactive page or notebook, a quiz, and an evidence list. It aims to fill the gap between demo videos, funding news, and scattered papers: from “understanding” to “doing” to “judging”, with every step checkable. This article covers who the course is for, how it is organized, and how far it has progressed.

Current status

These lessons are available in Chinese, English and Japanese. Videos and introductions are public; in-depth reading and exercises require sign-in. Other lessons are in production and review.

Physical AI Lab 0.1: What Is Physical AI? From Chatting to Doing · Watch video · Reading and exercises
Physical AI Lab 0.2: Why Now? The Three Curves of Data, Models, and Compute · Watch video · Reading and exercises
Physical AI Lab 0.3: See, Think, Act — Putting Each Rate Back in Its Own System · Watch video · Reading and exercises
Physical AI Lab 0.4: Ten Thousand Tries in a Virtual World — Putting the Conditions Back on Three Simulation Numbers · Watch video · Reading and exercises
Physical AI Lab 0.5: The Humanoid Robot Boom — Sorting One Passage of Tesla’s Quarterly Update into Demo, Prototype, Production and the Job · Watch video · Reading and exercises
Physical AI Lab 0.6: How to Judge Robot News — Tracing Each Question Back to a Sentence in the Source · Watch video · Reading and exercises
Physical AI Lab 1.1: Coordinate Frames and Kinematics — Working Out by Hand How a Two-Link Arm Reaches the Cup · Watch video · Reading and exercises
Physical AI Lab 1.2: PID and Control Rate — Computing the Three Terms by Hand, Then Turning Rates into Periods · Watch video · Reading and exercises
Physical AI Lab 1.3: Sensors — Working Out Pixels, Depth Range and Grip Force for One Cup · Watch video · Reading and exercises
Physical AI Lab 1.4: A Map of Simulators — Counting the Cost of the Timestep, Finding Where Contact Splits Two Runs, and Pinning Three Speed-ups to Their Baselines · Watch video · Reading and exercises
Physical AI Lab 1.5: Where Does Data Come From? Working Out Demonstration Time from SO-101 Metadata, Then Separating the Units and Action Spaces of Three Sources · Watch video · Reading and exercises
Physical AI Lab 1.6: Imitation Learning — Working Out How an Error Rolls Out of the Demonstrations’ Coverage, Then Putting the ACT, Diffusion Policy and Data Scaling Numbers Back in Their Own Conditions · Watch video · Reading and exercises
Physical AI Lab 1.7: Reinforcement Learning — Working Out One Episode’s Return to See How a Reward Gets Gamed, Then Reading the Paper’s Terrain Curriculum and Time-out Resets · Watch video · Reading and exercises
Physical AI Lab 1.8: Evaluation — Splitting a “Net Gain of 2” into a Paired Table, Then Checking the Failure Labels and Random Seeds · Watch video · Reading and exercises
Physical AI Lab 2.1: Building a Grasping Scene in MuJoCo — Working Out the Mass, Units, Timestep and Collision Rules the XML Leaves Unwritten · Watch video · Reading and exercises

Who it is for: five levels

The course is divided into five levels by audience; the higher the level, the more hands-on work and specialist background it requires:

Level	Audience	After completing it, you can	Episode format (planned)
L0 Understand	General public, media editors, investors	Explain what physical AI is and judge whether a robotics news story is credible	3–5 minute animated short + vertical clips
L1 Get started	Students, career changers, product managers	Understand the see-think-act loop and tune a parameter yourself in the browser	Explainer video + zero-install browser interactive
L2 Hands-on	Engineers	Train and evaluate a policy in simulation	Hands-on video + notebook (Colab or local GPU)
L3 Deep dive	Researchers, architects	Take apart VLAs, world models, and RL post-training; reproduce results and question them	In-depth lecture + reproduction experiments
L4 Industry	Decision-makers, trade media	Understand supply chains, costs, and compliance timelines, and judge deployment prospects	Explainer video + one-page brief

Engineers can start at L1 or L2, but the six L0 episodes are only a few minutes each, which makes them useful for aligning on concepts and for passing on to colleagues without a technical background.

Learning path: 36 episodes including capstone projects (planned)

The outline below is the course plan. Available lessons are listed above; the rest are in production and review. The plan may change with learner feedback.

L0 Understand (6 episodes)

0.1 What Is Physical AI? From Chatting to Doing
0.2 Why Now? The Three Curves of Data, Models, and Compute
0.3 See, Think, Act: One Second in the Life of a Robot
0.4 Practicing Ten Thousand Times in a Virtual World: What Simulation Is
0.5 The Humanoid Robot Boom: What Separates Demos, Prototypes, and Mass Production
0.6 How to Judge Whether a Robotics News Story Is Credible

L1 Get started (8 episodes, zero-install interactives)

Coordinate frames and kinematics; PID and control frequency; sensors (cameras, depth, force and touch); a map of simulators (MuJoCo, Isaac, Newton, Genesis); where data comes from (teleoperation, human video, synthetic data); imitation learning; reinforcement learning; and evaluation (what hides behind a single success rate).

L2 Hands-on (8 episodes, notebook exercises)

Building a grasping scene from scratch in MuJoCo; training ACT with LeRobot and SO-101 data; fine-tuning SmolVLA; training locomotion in parallel in Isaac Lab; domain randomization and system identification for sim2real; quantization, latency, and control frequency in edge deployment; control barrier function (CBF) safety filters; and reproducible evaluation with repeated trials and confidence intervals.

L3 Deep dive (6 episodes) and L4 Industry (4 episodes)

L3 covers taking apart VLA architectures, world-model data generation with Cosmos, RL post-training, long-horizon tasks and multi-policy routing, GR00T fine-tuning, and how to read papers and re-check their key numbers. L4 covers actuators and BOM cost, safety standards and regulatory timelines, which use cases will be deployed first, and finally a year-in-review.

L5 capstone projects (4 episodes, in production)

5.1 Kickoff: task and protocol
5.2 Data and baselines: scripted demos, imitation learning
5.3 High scores, so why does it break at a touch?
5.4 Passed the stress test. Ready to ship?

What each episode includes

An episode is more than a video. The planned deliverables are:

Main video: 16:9, with Chinese, English and Japanese narration and matching subtitles, plus vertical 9:16 material designed for mobile viewing.
Interactive: from L1 on, each episode comes with a zero-install browser page (TypeScript + three.js) or a runnable notebook. The same page that readers use is driven by a script and recorded into the video, so what the video demonstrates is the very page readers can open.
Quiz: 5 questions per episode, with explanations.
One-page brief: key points, key numbers with evidence levels, and further reading.
Evidence list and transcript: source links for every number and every claim in the episode, plus a transcript that can be turned into an illustrated article.

The structure of each episode is also fixed: a cold-open question within 8 seconds, a one-sentence promise, 2–4 chapters that each give an example before the concept, a three-point recap, and a question for the next episode. The series mascot, Zhixing (知行, roughly “knowing and doing”; UNIT-01), asks the questions and the follow-ups.

How “every number can be checked” is put into practice

This is the biggest difference between this course and typical popular-science videos, and it is written into the production rules:

Three evidence levels: every number and every claim is tagged with a level: O for official primary sources, R for academic literature (papers, preprints, academic monographs), and M for secondhand media reports. This matches the levels used by the Physical AI Radar.
The narration states what kind of source it is: an R-level number must be introduced with “the paper reports” (论文称 in the Chinese narration), and an M-level number with “according to reports” (据报道); only O-level numbers may be stated directly.
Every number gets a number card: whenever a number appears on screen, a card showing its level and source appears with it. Each piece of evidence carries a link and a date; arXiv papers use their v1 submission date.
No time words that go stale: the narration avoids words like “this week” or “today”, so statements don’t become wrong after the video is released.
Automated checks: after each build, a QA script opens every evidence link. Narration is checked against an automatic transcript, including proper nouns, loudness, peaks, silence and subtitle alignment. Visuals and content receive an independent review.
Course as code: scripts, shots, and numbers all live in source files. Any episode can be regenerated from source, and changing a number lets you trace which shots it affects.

Before release there is also a 10-item review scorecard worth 20 points in total; the bar is at least 17 points with no item scored 0. Episodes at L2 and above also get their technical content reviewed by peers in the field.

How to follow along

The Courses section of this site will publish a text version of each episode that explains the concepts, lists the numbers and how to check them, and includes self-check questions. Once a video goes live, a link will be added to the corresponding article.
Each episode’s interactive pages and notebooks will be released with that episode, and their addresses will be given in the article at that time.
To get familiar with the course’s evidence conventions in advance, see the Physical AI Radar: it organizes physical AI developments along eight lanes every day and uses the same O/R/M levels.

Limitations

The outline is a draft; the number of episodes, titles, and order may all change. Planned items listed here do not mean they have been released.
“Checkable” means the original source can be found; it does not mean the original source is necessarily correct. R-level sources are mostly preprints that have not been peer-reviewed, and O-level sources are often self-descriptions by vendors or institutions; the course treats numbers from both as “reported by the authors”.