Where you are. Seven modules are behind you: the loop, the maths, the simulator, trained policies, a real arm with your own dataset on it, a fine-tuned vision-language-action model, and an agent planning over learned skills. This lesson spends that history on one decision.
The commits nobody asked you to make
Open your course repository and look for the work that was not on the plan.
Somewhere in the last nine months there is a weekend where you were meant to be doing one thing and did another. An eval harness that grew three features nobody needed. A tuning run you kept extending past the point of usefulness. A bug you chased for six hours after the deadline had already passed, because not knowing had become intolerable.
Now find the opposite. The piece of work you completed exactly to spec, on time, and have never opened since.
Nobody assigned either behaviour. That difference is the only honest data you own about your own motivation, because it was not collected while you were trying to impress anyone, including yourself.
It is also not sufficient. Motivation on its own picks badly, and the rest of this lesson is the correction.
The idea in one paragraph
Specializing is not choosing the field with the best headlines. It is choosing which of your existing advantages you are going to compound, then committing enough calendar time that the choice becomes visible to other people. Four things decide it: pull, meaning what you do when nobody is watching; compounding, meaning where a decade of distributed-systems work is worth years here rather than weeks; burn, meaning what the track costs in money, compute and calendar before it produces anything you can show; and market shape, meaning whether a self-directed year clears the bar for the roles the track leads to. Weigh them in that order, with one correction applied to the first: novelty feels exactly like interest for about three weeks, and three weeks is shorter than any track.
Pull is evidence, and it is noisy
Pull is the strongest single predictor of whether you finish a capstone, because a capstone is six to eight weeks long and no external pressure survives that. It is also the criterion most easily faked by a good demo video.
The test that separates real pull from novelty is return. Not “did I enjoy it”, which is unreliable, but “did I come back to it after the new-thing feeling wore off, without a reason”. Two separate returns, at least a month apart, is a signal. One enthusiastic weekend is not.
Compounding is the criterion people discount because it feels like cheating
You are not paid for total skill. You are paid for the scarcest combination of skills you can hold at once, and combinations are much rarer than components.
That has a sharp consequence here. There are thousands of people with more robotics than you and thousands with comparable distributed-systems depth. The number holding both is small, and the number holding both plus real fluency in agent and tool-use architecture is smaller still. A track that uses one of your existing strengths is pleasant. A track that requires two of them is a position.
Wider than the screen; scroll it sideways.
Place your own three dots on that grid before you read anyone else’s opinion, including this course’s. The top-right quadrant is the only one that survives a bad month.
Burn: what the track costs before it shows anything
Every track eventually produces an artifact. They differ enormously in how much you spend first, and the spend is not mostly money.
Wider than the screen; scroll it sideways.
Calendar-to-first-artifact is the one to watch, because it decides how many times you get to be wrong. A track where you can produce something showable in two weeks lets you correct course four times inside a capstone. A track where the first credible result takes two months lets you correct once, and only if nothing breaks.
Market shape, and why salary tables are the worst available input
There is a large volume of robotics compensation content published every year, and almost all of it is recruiting collateral with an incentive to quote the top of the band. Two recruiter-published tables I compared in August 2026 disagreed with each other by 20 to 40% at senior level for the same titles. That disagreement is the finding. It tells you the underlying data is thin, self-reported, or both.
Here is what survives a source check, with dates attached so you can tell when it rots.
| Source | What it actually says | Caveats |
|---|---|---|
| NVIDIA job posting, Senior Robotics Systems SWE (ROS), Santa Clara, 8+ years | Base $184,000-$287,500 at one level, $224,000-$356,500 at the next | Employer-published, so credible. Posting removed November 2025, so around nine months stale. Base only; excludes substantial equity |
| levels.fyi, Boston Dynamics software engineer, read 9 August 2026 | Median total compensation around $220K; highest reported $307,166 | Self-reported and a small sample, which levels.fyi shows on the page |
| Two recruiter salary guides, first half of 2026 | Senior bands from roughly $195K to $345K depending on title | Undisclosed methodology, inflation incentive, and they contradict each other |
Use market shape differently. Ask which roles a track leads to, how thin the category is, and whether the gate is a credential you cannot manufacture. Thin categories are good for you: fewer competitors, and hiring managers who cannot easily find anyone. Gated categories are bad for you regardless of interest.
The three tracks, stated plainly
The next three lessons take one each. In summary:
| Track | What it is | Compounds with | The honest catch |
|---|---|---|---|
| Agentic robotics | A planner over learned skills, hardened: multi-step tasks, verification between steps, an eval suite, possibly a second embodiment | Agent architecture, tool-use design, distributed systems | A thin category. It must be your second exhibit, not your first, or it reads as avoiding the robotics |
| Locomotion and humanoids | Reinforcement learning in simulation, curriculum and domain randomisation, sim-to-real onto legged hardware | Little of your background; this one you buy with time | Highest burn on every axis, and the roles it targets have the hardest credential gate |
| Data and fleet infrastructure | Collection tooling, dataset schemas and validation, eval farms, deployment and fleet observability | Nearly all of your background, almost directly | Least glamorous. Which is also why it is underpriced |
The recommendation this course makes, and it is a recommendation rather than a verdict, is agentic robotics as the primary track with data infrastructure as the fallback if the pull is not there. Both use the background you actually have. Locomotion is the right answer only if legged robots genuinely pull you, in the returning sense above, because nothing else about it is easy.
What the choice actually commits
Less than it feels like.
Wider than the screen; scroll it sideways.
Everything before the fork is shared: the honest ROS 2 minimum, logging and visualisation fluency, and the writing habit. Those three are in this module for every track, and they are portable to any robotics job. What the track choice buys is the shape of one capstone and the sentence you lead with when someone asks what you do.
Check yourself
1. Why is “what pulls me” ranked first but corrected immediately?
Because a capstone is six to eight weeks long and no external motivation survives that distance, so pull is the best predictor of finishing. The correction is that novelty is indistinguishable from interest for roughly three weeks, which is shorter than any track. The test that separates them is return: did you come back to this after the new-thing feeling wore off, without a reason, at least twice, at least a month apart.
2. You are not paid for total skill. What are you paid for, and how does that change the ranking of the three tracks?
For the scarcest combination you can hold at once. Combinations are much rarer than components, so a track that requires two of your existing strengths is worth more than a track that uses one well. That pushes agentic robotics and data infrastructure up, because both need a background you already have plus the robotics you just acquired, and pushes locomotion down, since almost none of your prior work compounds inside it.
3. Two published salary guides disagree by 20 to 40% on the same senior title. What should you conclude?
That the underlying data is thin, self-reported, or produced by someone with an incentive to quote high, and that neither number should anchor anything. Use them only as a very wide range. The claims that survive a check are employer-published postings, with their date noted because they go stale in months, and self-reported aggregates with the sample size visible. Neither of those describes what you personally will be offered after one self-directed year.
4. Why is “calendar to first artifact” a better burn measure than money?
Because it sets how many times you are allowed to be wrong. Money is a one-time gate you either clear or do not. Calendar is a rate: a track that shows something in two weeks lets you correct course several times inside a capstone, while a track whose first credible result takes two months lets you correct once, assuming nothing breaks. Since your track choice is a hypothesis rather than a certainty, the cost of testing it matters more than the cost of entering it.
5. What does the track choice actually commit you to, and what does it not?
It commits the shape of one capstone, six to eight weeks, and the sentence you lead with when someone asks what you work on. It does not commit the systems literacy, the logging and visualisation fluency, or the writing habit, because those are shared across all three tracks and portable to any robotics role. The genuinely irreversible cost is the opposite of what people fear: not choosing wrong, but choosing two and shipping neither.
Do this
About ninety minutes, and it produces a document you will reread at the end of the module.
1. Run the evidence audit. Go back through nine months of commits, notes and reading history. Write down every occasion where you did unassigned work, with a one-line description and a date. Then mark which of the three tracks each one belongs to. You are looking for returns, so ignore anything that happened once.
2. Place three dots on the grid. Using the pull-versus-compounding figure, place each track. Be specific about the compounding axis: name the skill from your existing background that transfers, and name the thing it lets you skip. If you cannot name one, the dot is further left than you drew it.
3. Cost the burn. For each track, write your real numbers: hardware you would need to buy, GPU hours or cloud spend per month, and honestly, how many weeks until you have something a stranger could look at.
4. Write notes/07-track-memo.md. One page. Which track, why, in terms of pull and compounding rather than interest. What you are explicitly giving up. One condition that would make you change your mind, written as something checkable rather than a feeling, such as “if I have not produced a runnable artifact in three weeks”. Date it.
Then read the three track lessons, in order, even for the tracks you did not pick. Each one names failure modes that generalise.
What you can now do
You can rank the three tracks on four criteria instead of on how interesting they sound, tell real pull from novelty using return rather than enthusiasm, name the specific place your existing background is mispriced, read a robotics compensation claim sceptically enough to notice when two sources contradict each other, and state in one page what you are choosing and what you are giving up to choose it.