Where you are. The capstone is built, shipped and written up. This is the last lesson of the course, and it has one job: turn twelve months of work into something other people can evaluate without you standing next to it.
Open the note you wrote in week one
notes/00-thesis.md. One page, written in the first week of Module 0, before you had touched a servo. Open it and read it straight through without editing anything.
Three kinds of sentence are in there.
Some you would write again today, more or less unchanged. Those were not luck. They were the parts you had actually reasoned about rather than absorbed from someone else’s writing.
Some are wrong in an interesting way, and you can name the moment they broke. A generalisation claim that survived until the eleventh object. A number from a paper your own evaluation harness would not reproduce.
And some are wrong in a boring way: fluent, confident, assembled from other people’s summaries of a part of the field you had never touched.
You can sort that page into three piles in fifteen minutes, and the sorting is the most accurate measurement of the year that exists anywhere. It is also completely invisible to everybody except you. It does not appear on a profile, a résumé or a call. The rest of this lesson is about the part that does.
The idea in one paragraph
A position is not a skill list. It is a small set of objects a stranger can check without you present, plus one sentence saying which problem you own. Four things build it, in this order: artifacts, because everything else refers back to them; recall under pressure, because an interview is mostly a test of whether the artifacts were really yours; a few real relationships, because thin categories hire through people rather than through job boards; and a claim, meaning a dated bet about where the field goes next, which is what the rewritten thesis note becomes. The finish line is not a feeling of being ready. It is a checklist where every item is either present at a URL or absent.
Everything gets read without you
Assume seven minutes and no follow-up questions. Someone opens your profile the evening before a call, or between two other candidates, and forms a view. They will not ask you what you meant.
Wider than the screen; scroll it sideways.
The last box is the real output. Nobody forwards a repository; they forward a sentence about a repository. Every object upstream exists to make that sentence specific rather than generic.
This has an unpopular consequence. One artifact that answers a question completely beats five that each answer part of one, because the five are read in the same seven minutes and none of them finishes a thought.
What the portfolio has to contain
In priority order. This is the checklist the course has been building toward since Module 1, and most of the items already exist somewhere in your repository; the work now is finishing them and making them findable.
| Object | What it answers for a stranger | Done when |
|---|---|---|
| One end-to-end system | can this person build something that touches the physical world and survives it | perception or teleoperation, through a policy or planner, to a real actuator, running, with a limitations section that is longer than one line |
| Uncut video, failures included | is any of this real | at least one take that is not the good take, on hardware, unedited |
| An evaluation harness with numbers | do they know what “working” means | success rate over a stated number of trials, broken down per condition, with the protocol written down |
| A published dataset | can they produce data other people can use | episodes in a standard schema on a public hub, with a card saying how it was collected and what is wrong with it |
| A replayable recording | tooling literacy | a bag file plus a layout a stranger can open and scrub through unaided |
| A written post-mortem | can they debug | one failure traced to a root cause, with the fix and what it cost |
| The agentic exhibit | the thing almost nobody else has | a planner over a real robot interface, with honest latency and reliability numbers |
Two notes on ordering. The dataset ranks that high because it is the single most credible object for the infrastructure roles, and because a dataset is used by other people, which no video ever is. The agentic exhibit ranks last deliberately: leading with it reads as avoiding the robotics, and the track lesson says why at length.
What you should be whiteboard-ready on
Short list. Each item is one you can produce standing up, with a marker, being interrupted.
- Frames and transforms, forward and inverse kinematics, and why composition order matters.
- Imitation-learning architectures against each other: what an action-chunking policy buys, what a diffusion policy buys, what a vision-language-action model buys, and what each costs in data and latency.
- The reality gap: where it comes from, and the three specific things you did about it.
- Data-engine design: how you would collect ten times the data with the same number of humans.
- Evaluation methodology: how you would know a policy got better, and what “better” means when the variance across trials is large.
- Agentic architecture: where verification sits, and why it belongs between skills rather than inside them.
- The industry-interface minimum from the ROS 2 lesson: a quality-of-service mismatch dropping messages silently, and the deadlock you get calling a service from inside a callback on a single-threaded executor.
None of that is exotic. The difficulty is not recall, it is producing it live, and the reason interviews work that way is the depth probe.
Wider than the screen; scroll it sideways.
Level one is answerable from a README. Level two needs a decision you made, and level three needs the consequence of that decision measured. Preparation cannot fake this, which is exactly why it survives as a filter, and it is the structural reason one genuinely hard debugged system beats a portfolio of tutorials.
The network is five people, not five thousand
Thin categories do not hire through job boards, because the posting attracts hundreds of people who match on keywords and nobody who matches on the combination. They hire through someone saying a name.
You are not building an audience. You are building three to five genuine relationships with people whose work you have actually used, which happens through a small number of unglamorous acts: answering beginner questions in the community where you learned, publishing something that saves someone else a weekend, and getting one contribution merged into a tool people depend on.
The roles this unlocks, ranked by fit rather than glamour
| Role | Why it fits you | The gate |
|---|---|---|
| Robotics data and infrastructure engineer | dataset pipelines, ingestion, training infrastructure, evaluation harnesses; a near-direct port of a distributed-systems background | a dataset and an evaluation harness other people can run |
| Simulation and evaluation engineer | simulators, randomisation, continuous evaluation for policies; undersupplied and mostly a software problem | evidence you have measured a policy honestly, not just trained one |
| Agentic robotics and robot-facing tooling | planners over robot stacks, tool interfaces, teleoperation and fleet interfaces; your rarest combination | robotics depth demonstrated first |
| Developer relations or solutions architecture at a robotics tooling company | prices in the role you already hold instead of discounting it to zero | public writing and talks, which this course has been producing since Module 1 |
| Robot learning engineer | the title with the most attention on it | usually hands-on policy work on real hardware at a recognisable programme; a self-directed year is filtered out more often than not |
| Robotics platform engineer | highest density of the industry-standard middleware | production C++, real-time Linux, middleware debugging |
| Forward-deployed engineer at a humanoid or warehouse company | integration into messy real facilities, growing as pilots convert to production | travel, and owning an unglamorous problem end to end |
Two honest notes. Production C++ is a real gate, not a preference: platform postings say C++17 or C++20 and mean it. Being Python-only closes the platform door and leaves the learning, data and tooling doors wide open, which is fine, but choose it knowingly. And on compensation, the track lesson has the sourced numbers; the summary is that a senior software background plus a self-directed year prices as “senior engineer who happens to work on robots”, robotics generally pays below big-tech backend at equivalent seniority, and the upside arrives as equity, which is an option rather than cash.
Rewrite the note
Now go back to that page from week one and write today’s version next to it.
Wider than the screen; scroll it sideways.
Keep both files. The week-one version is evidence, and evidence you have edited is worth nothing. The new one has three sections and takes a page: what held, what broke and which experiment broke it, and what you are betting on next. Then a date, and a review date roughly a quarter out.
The bet is the part people skip and the part that matters. A position needs a claim in it, because “I have these skills” is a description and “I think the field goes this way and I have built the thing that pays off if it does” is a stance. Someone can disagree with a stance, which is the point: it is the sentence that makes a conversation start rather than end.
Check yourself
1. Why is “delete yourself from the room” the test for every portfolio object, rather than “is this impressive”?
Because the read that decides things happens without you: seven minutes, before a call or between two other candidates, with no follow-up questions available. Anything that needed your narration to be legible converts to nothing in that read. It also changes what you build. A demo you can talk over becomes a demo with a written protocol, a numbers table and a limitations section, because those are the parts that speak on their own.
2. Two candidates. One has five tutorial repositories, the other has one system with an uncut failure video and a success-rate table. Name the specific mechanism the second survives and the first does not.
The depth probe. The interviewer asks which problem consumed most of your time and follows up three levels down. Level one is answerable from any README. Level two needs a decision you personally made, and level three needs the measured consequence of that decision. Tutorial work has no level two, because the decisions were made by whoever wrote the tutorial. This is also why building two half-capstones is worse than one whole one: breadth adds level-one answers and no depth.
3. Why does a published dataset rank above a trained checkpoint in the portfolio list?
Because other people can use it, and use is the strongest signal available. A checkpoint is evidence you ran a training job; a dataset in a standard schema, with a card saying how it was collected and what is wrong with it, is evidence you can specify a schema, run a collection protocol, find your own defects and ship something with a contract other people depend on. That is precisely the job description for the infrastructure roles, which are the best-fit roles on the list.
4. The course ranks robot learning engineer sixth on fit while calling it the title with the most attention. Is that a contradiction?
No, because fit is about the gate rather than the appeal. That role is usually gated on hands-on policy work on real hardware inside a recognisable programme, which is a credential a self-directed year does not manufacture, so applications get filtered before anyone reads the artifacts. The higher-ranked roles are gated on things you can produce alone and show: a dataset, an evaluation harness, a running system, public writing. Rank by which gate you can actually clear this year, then move sideways later from inside the industry.
5. Nobody else will ever read your thesis note. Why rewrite it, and why does it need a date?
Because a position needs a claim in it, and the note is where you make the claim explicit enough to check. “I have these skills” is a description and invites nothing. “The field goes this way, and here is the thing I built that pays off if it does” is a stance someone can argue with, which is how conversations start. The date and the review date turn it from an opinion into something falsifiable: in three months you can see which parts held, which broke, and what broke them, and that is the same delta you just measured across the year.
Do this
Half a day, and it closes the course.
1. Run the seven-minute read on yourself. Open your public profile in a private browser window, set a timer, and read it as a stranger who has never met you. Write down what you learned, and only what the page actually told you. Then write the gap list.
2. Fill the checklist with URLs. For each of the seven objects in the table, write the link or write “missing”. No partial credit and no “mostly done”. Anything missing that can be finished in a day gets finished this week; anything that cannot gets a date.
3. Rehearse three whiteboard topics. Pick the three you feel least solid on. Record yourself for five minutes each, no notes, standing up, and watch it back. Every level-two follow-up you could not have answered goes on a study list, and the list is usually short.
4. Write the one sentence. Fifteen words, no adjectives, of the form “I build X for Y”. Test it on someone outside robotics. If they cannot repeat it back, it is not the sentence; the failure is almost always that it names a technology instead of a problem.
5. Rewrite the thesis note. Copy the original to notes/00-thesis-v1.md so the week-one version survives unedited, then rewrite notes/00-thesis.md in three sections: held, broken with the experiment that broke it, and what you are betting on next. Date it. Put a review date a quarter out in your calendar.
6. Send three messages. To three people whose work you actually used this year. One paragraph each: what you built with it, one specific thing you learned that they might not know, and no ask attached. The ask comes later or never, and either is fine.
What you can now do
You can audit your own work the way a stranger reads it, name the seven objects a credible robotics portfolio contains and which of them you are missing, explain why an evaluation harness and a published dataset buy more credibility than another demo, hold your own through three levels of a depth probe on the system you built, rank the roles this year unlocks by which gate you can actually clear, and state a dated bet about where the field goes next.
That was the last lesson. What you should carry out of it is smaller than the syllabus and older than any of the tools: you have spent a year running the same loop the robot runs. Sense what is actually true rather than what you assumed, decide with the information you have, act, then go and find out what really happened. Every module was that loop pointed at a different subject, and the thesis note you just rewrote is one more turn of it, on the largest timescale in the course.
The course ends here. The loop does not.