Physical AI
Syllabus · Draft course plan
| Textbook | Volume IV: Physical AI, a working draft |
| Prerequisites | Python, introductory machine learning, basic linear algebra and probability, and first-year Newtonian physics |
| Format | A 14-week semester or a 10-week quarter, with a 90-minute seminar and a 3-hour supervised lab each week |
| Hardware | One robot station per team: a Seeed SO-101 arm, an overhead camera, and an Arduino UNO Q (station reference; draft parts list with planning prices) |
| Materials | Slides · Lab sequence · ETH Zurich spring 2027 syllabus |
Description
Students build and evaluate a robot whose learned policy proposes each action and whose independent governor decides whether to permit it. Students then measure the resulting motion, and a new observation decides what happens next. The course is a proposed hardware studio and has not yet been taught. These materials are still in preparation for teaching in the next semester or year, so they will change as the course takes shape.
Learning outcomes
By the end of the course, students can:
- Trace the path from sensing to actuation and name the component that proposes, permits, and executes each action.
- Calibrate sensors, measure actuator limits and latency, and set an operating envelope from those measurements.
- Compare a simulation with the real hardware and explain where they disagree.
- Train a learned policy, deploy it on edge hardware, and compare it with a nonlearned baseline.
- Build and test a safety governor, and defend a release decision with evidence from their experiments.
Reading schedule
In the semester plan, students read one or two Volume IV chapters each week. The lab sequence pairs each week with work at the robot station.
| Week | Reading | Slides |
|---|---|---|
| 1 | The Causal Boundary | Ch 1 |
| 2 | The Physical Body | Ch 2 |
| 3 | The Cognitive Brain | Ch 3 |
| 4 | The Nervous System | Ch 4 |
| 5 | Physical Data | Ch 5 |
| 6 | Policy Training; read ahead: calibration in Ch 8 | Ch 6 |
| 7 | Closed-Loop Evaluation | Ch 7 |
| 8 | Sensor Perception, Spatial Memory | Ch 8 · Ch 9 |
| 9 | Grounded Intent, Trajectory Planning | Ch 10 · Ch 11 |
| 10 | Safety Enforcement, Silicon Placement | Ch 12 · Ch 13 |
| 11 | Supervisory Intervention, Adversarial Verification | Ch 14 · Ch 15 |
| 12 | Safe Release | Ch 16 |
| 13 | Responsible Physical AI | Ch 17 |
| 14 | The Epistemic Frontier | Ch 18 |
ETH weeks 2–3 and 5–6 are combined, and the last week covers release, responsibility, and the frontier together. Staff qualification of the station takes the same time in either format.
| Week | Reading | Slides |
|---|---|---|
| 1 | The Causal Boundary | Ch 1 |
| 2 | The Physical Body · The Cognitive Brain | Ch 2 · Ch 3 |
| 3 | The Nervous System | Ch 4 |
| 4 | Physical Data · Policy Training | Ch 5 · Ch 6 |
| 5 | Closed-Loop Evaluation | Ch 7 |
| 6 | Sensor Perception, Spatial Memory | Ch 8 · Ch 9 |
| 7 | Grounded Intent, Trajectory Planning | Ch 10 · Ch 11 |
| 8 | Safety Enforcement, Silicon Placement | Ch 12 · Ch 13 |
| 9 | Supervisory Intervention, Adversarial Verification | Ch 14 · Ch 15 |
| 10 | Safe Release · Responsible Physical AI · The Epistemic Frontier | Ch 16 · Ch 17 · Ch 18 |
Assessment
The ETH Zurich plan uses this split; each offering sets its own.
| Component | Weight |
|---|---|
| Labs 1–2: the station and its operating envelope | 15% |
| Labs 3–4: a learned policy on the edge | 25% |
| Lab 5: a stress-tested loop | 10% |
| Lab 6: a safety governor | 25% |
| Capstone: a release decision defended with evidence | 25% |
Before teaching
- The station is a proposal. Before students use it, staff must show the complete command path working on the assembled hardware, including the independent motor cutoff, the operating limits, and restart behavior. Connecting a host directly to the arm’s motor bus does not by itself give the microcontroller authority over each command.
- Training needs workstation compute. The course has not fixed a minimum training configuration, policy, simulator, or export path. Staff choose these and measure memory use and run time on the actual equipment. The introductory course’s no-GPU assumption does not apply here.
- A trial run comes first. Staff run the full student workflow once before setting enrollment or buying a class set. The trial must show the policy proposing an action, the governor permitting or refusing it, and the arm responding within the measured limits. It also sets station capacity, supervision, and bench time.
The hardware feasibility plan contains the detailed staff preparation. It is a planning document and does not establish that the station has passed qualification.
Adapting this syllabus
- Each offering sets its own dates, deadlines, and policies on late work, collaboration, and AI tools.
- An offering can change the task, term length, and hardware once staff qualify the resulting station. The system students build must still separate the policy’s proposals from the governor’s permission.
- The ETH Zurich spring 2027 syllabus gives one complete offering, with its labs and assessment.