Volume IV: Physical AI Studio Labs
These reusable labs accompany MLSysBook Volume IV. Teams measure a robot’s physical limits, compare simulation with real hardware, collect demonstrations, deploy a learned policy, and test an independent governor.
Status: Lab skeletons; staff are preparing and qualifying the instructions.
The course overview, semester schedule, and assessment policies now live on the Physical AI course website. See the course syllabus for readings and deadlines.
Lab Instructions
- Getting Started: Prepare the station and learn the operating rules.
- Lab 1: The Digital Twin: Measure the body and build a kinematic model.
- Lab 2: Simulation to Real Hardware: Compare predicted and measured motion and timing.
- Lab 3: Learned Control: Compare a learned policy with the nonlearned baseline.
- Lab 4: Real Data and Fine-Tuning: Collect demonstrations, adapt the policy, and evaluate the loop.
- Lab 5: Stress-Testing: Test action horizons and disturbance recovery.
- Lab 6: Safety Governor: Build the governor and test its declared fault responses.
- Capstone Studio: Defend a release decision with physical evidence.
Chapter companions: explore before the physical studio
The chapter-aligned browser companions complement this hardware studio. Start with lab_00_orientation.py, then follow the chapter order through the final transfer investigation. Each chapter uses four concept-owned parts, visible goals and route, reactive reference evidence, saved comparisons, and one downloadable discussion. The chapter inventory connects every companion to the studio session it supports.
The browser project is the book’s warehouse mobile manipulator. The studio uses the SO-101 station. A warehouse reference calculation is not a measurement of that station, and a constructed incident trace is not a physical trial. These companions issue no hardware commands and do not alter the studio’s status or replace staff qualification. Physical instructions and approved station limits remain in the existing studio and Station Reference Card.
Concepts and student-visible goals are authored once in concepts/; the packaged learning contracts are generated from those maps. MLSysIM owns study cases and calculations. mlsysbook_labs.physical_companion owns the shared interaction; build_companions.py checks the thin chapter entrypoints. No chapter notebook contains a private copy of the physics or narrative.