MODULE 30 / PHYSICAL WORKSHOP
Close the loop with the real machine
Finish with an independent investigation and a route to physical iteration. Learn system identification and regression testing through a reversible model change.
By the end, you can…
- Design an experiment separating possible causes
- Update a model consistently
- Prepare a reproducible candidate and physical follow-up
All previous modules
A reproducible baseline
Physical robot
Measure lengths, mass, joint axes, mounting offsets and materials. Record uncertainty.
The digital robot is a hypothesis
The model predicts what the robot should do under specified conditions. Hardware measurements eventually reveal where that description is wrong. System identification uses measured input/output behavior to estimate properties such as inertia, friction, delay or motor response. It is a controlled comparison, not simply changing parameters until an animation looks right.
Start with the simplest observable mismatch. Wrong joint direction suggests conventions or mapping. Delayed response calls for timestamp and transport measurements before assuming incorrect mass. Slower acceleration after adding weight calls for mass/inertia checks. Several parameter combinations can resemble one another in one trace; additional experiments may be needed to distinguish causes.
Make consistent physical changes
A payload changes mass, inertia and center of mass according to its location. Adding 100 g at the head differs from adding it near the torso. If a body has explicit inertia, changing only mass can create an inconsistent model. Preserve a clear physical description of the change.
The teaching hinge derives inertia from geometry and mass. Changing its link mass in a working copy gives a controlled example. Keep geometry, gains, time step and target unchanged. Predict the effect, measure it and restore the reference. Later hardware iterations follow the same process, with measured quantities replacing invented values.
Protect behavior with regression checks
A candidate should pass named checks: startup, standing, forward, backward, turning, transitions and selected perturbations. Define metrics and comparison intervals before seeing results. Preserve failures. Better tracking in one command can coexist with worse stopping, and a small average error can hide a brief excessive tilt.
Choose thresholds appropriate to the experiment. Deterministic local runs may compare tightly; hardware has sensor noise, battery variation and surface differences. Record conditions, repeat trials and report ranges where appropriate. The course’s 60-degree numeric termination is a simulation escape condition, not an acceptable physical tilt specification.
Make a release another person can reproduce
A release record names source commits, policy/model hashes, dependencies, calibration version, component configuration, test results and limits. Keep the last accepted version available. An edited battery position and a new policy are a different candidate even if both directories are named final.
Your capstone demonstrates judgment: what changed, why it is physically meaningful, what the data support and what remains open. Course completion gives you a method for manufacturing and commissioning. The exact bill of materials, reinforced geometry and electronics compatibility still require their own verified decisions.
WORDS YOU NOW KNOW
- System identification
- Estimating model properties from measured inputs and responses.
- Regression suite
- Repeated checks that detect lost behavior after a change.
- Release record
- The exact configuration, evidence and limits attached to a candidate.
YOUR CALL
Finished with this module?
0 of 8 lab steps checked. 0 of 3 questions answered. You decide when the module is complete. Revisiting or editing it will not change your decision.
You can reopen it later. Checking every box will not automatically complete it.Go to the source
Primary references for this module. Use the version and hardware revision named in the lesson.
MuJoCo model and inertia conceptsBDX-R source repositoriesKayden’s system identification workDownload lab setup notes