Projects, portfolio and capstone assessment
The portfolio is an evidence trail, not a gallery of unexplained success videos. Keep code, data and notebook entries in your own private repository. This site collects no student work. Any public sharing is optional and should be reviewed for personal information and permissions.
Six phase checkpoints
| Phase | Required evidence | Explain and defend |
|---|---|---|
| 1: programming/electronics/motion | system map, explained Python sequence, LED measurements, bounded pulse trials | what the command does and what it cannot measure |
| 2: measurement/kinematics | raw CSV, counts-distance calibration, turn model, independent validation | units, residuals, uncertainty and operating conditions |
| 3: feedback/PID | baseline comparison, gain sweep, bounded P/PI/PID tests, held-out result | why a term helps or hurts under these conditions |
| 4: sensing/avoidance | labelled sensor data, confusion counts, state diagram, fault matrix | false stops, missed hazards and incomplete sensing |
| 5: localization/navigation | odometry trace, landmark gate tests, inflated map, BFS fixtures, route trials | drift, association, map assumptions and endpoint truth |
| 6: capstone | requirements, integration baseline, full validation batch, report and defence | the measured claim, failures and next experiment |
For each checkpoint, a well-documented failed prediction can satisfy reasoning goals. Physical completion remains separate from software/replay completion: explicitly mark which evidence types you have.
Example report — fictional measurement study
Question: does a start guide reduce variability in a fixed pulse? Prediction: guided starts have smaller distance spread. Procedure: same command, 0.5 s duration, mat and measuring method; five trials per condition, alternated when possible. Criterion: retain the guide if spread decreases without a large unintended shift in mean. Define “large” before the actual test.
| Condition | Illustrative distances, cm | Mean, cm | Range, cm |
|---|---|---|---|
| Original | 18, 20, 21, 19, 22 | 20 | 4 |
| Guided | 19, 20, 20, 21, 20 | 20 | 2 |
Observation: in these fictional values, the mean stays 20 cm and range falls from 4 to 2 cm. Interpretation: more consistent starting alignment may explain the reduced spread. Alternative: operator practice or battery changes may also matter. Uncertainty: small batch, ambiguous centre measurement and one surface. Decision: the illustrated result supports a larger alternated validation batch, not a universal accuracy claim. An actual report links its raw CSV, plotted points and source revision.
Capstone mission options
Core physical mission: navigate a short surveyed clear route on a level mat using encoders and a planned waypoint list; stop for contact/boundary, invalid data or timeout. Measure endpoint independently. No unknown-environment non-contact avoidance is promised.
Software mission: use the desktop lab kit and your own planner/controller variant; test normal, blocked-route, stale, hazard and stalled cases. Label every result synthetic. A complete software portfolio does not claim physical validation.
Optional extension: add a qualified sensor or automatic landmark correction only after the baseline, electrical compatibility and new failure tests are complete. Preserve the original baseline so you can compare evidence.
Suggested mission requirements
Adjust these before testing to match your measured capabilities: arrival within 0.10 m of a surveyed goal, completion within 60 s, no boundary breach, and deterministic stop on every injected invalid/hazard case. These are teaching examples, not manufacturer guarantees. Define route/start/surface, map clearance and measurement method. Freeze the rules before validation.
Use at least ten held-out navigation attempts across two declared conditions where feasible, retaining every outcome. If physical access or time limits the batch, report the actual smaller count and narrower evidence. Separately evaluate software guard cases. Safe stopping on a hazard supports the guard requirement while the same trial can fail navigation.
Assessment rubric
Score each dimension from 0 to 3 using the descriptions below. The total is a discussion aid, not proof of physical safety. No unsupported claim or missing evidence is repaired by a high score elsewhere.
| Dimension | 0: missing | 1: emerging | 2: clear | 3: defensible |
|---|---|---|---|---|
| Requirements and design | no measurable mission | vague target or unsupported sensing | measurable scoped mission and interface diagram | justified trade-offs, assumptions and independent tests |
| Code and integration | unexplained/nonreproducible | partial integration or guessed API | readable modular code, calibrated units and guards | replayable decisions, fault tests and exact version/configuration |
| Experiment quality | no retained data | selected successes or weak controls | full raw outcomes and fair comparison | held-out/version-separated tests with confounding/uncertainty discussed |
| Reasoning and math | unsupported assertions | arithmetic without assumptions | correct units, models and interpretation | independent checks, limitations and alternative explanations |
| Communication and handoff | cannot locate work | incomplete README/plots | linked report, labelled plots and setup | independent replay/check and evidence-led defence |
Physical validation status is recorded separately: not attempted, pilot only, or qualified within stated conditions. An unpiloted program is never labelled hardware validated.
Final portfolio checklist
- README with purpose, exact robot/firmware, setup, directory map and run instructions.
- Requirements and surveyed map with dimensions, coordinate convention and clearance.
- Source/configuration and final tested Git tag/revision.
- Raw trial records, including stopped failures and missing-measurement reasons.
- Analysis code, labelled plots and version-separated summary metrics.
- Notebook entries and design/decision history with sources and AI assistance.
- Five-minute explanation/demo, independent endpoint measurement or labelled simulation truth.
- Limitations, unresolved faults and three ranked next experiments.
Five-minute engineering defence
Spend about one minute on the challenge/requirements, one on the model/design, one on the experiment, one on results including failures, and one on limitations/next steps. Answer: What evidence says feedback helped? How could a wrong landmark hurt? Which new condition invalidates your claim? Where is the raw record behind your best plot?
A useful final claim names version, attempts, conditions, errors and stops. For example, “This fictional version arrived in 8/10 surveyed-mat trials; arrived error averaged 6 cm; two missions stopped on hazards.” Avoid “always works.” Invite a peer to reproduce one calculation or replay, and archive the final baseline before beginning an extension.