A software engineer gets into robotics and physical AI by adding three things to the code they already write: perception (making sense of cameras and depth sensors), control (making a machine move where you mean it to, safely), and the discipline of testing in simulation and then on real hardware. You do not need to become a mechanical engineer, but you do need to be comfortable with the physical world being messier than any test suite. The job ads say the same in their own way: in 33 postings we counted at robotics and physical-AI employers, 58% said on-site and 42% mentioned a degree.
What do robotics job ads ask for?
We re-ran our job-ad sample on 29 September 2026: 3,610 public postings from Hacker News "Who is hiring?" (July to September 2026), Remotive and Arbeitnow. 33 were at robotics or physical-AI employers, 13 from Hacker News and 20 from Arbeitnow. Read the table with one caveat in mind: those 33 include mechanical, electrical, marketing and full-stack roles at robotics companies, not only robotics software engineers.
| What the ad mentions | Robotics employers (33) | All AI ads (315, 19 Sep) |
|---|---|---|
| On-site | 58% | 32% |
| A degree | 42% | 12% |
| Python | 30% | 38% |
| C++ | 27% | 4% |
| A PhD | 18% | 3% |
Three things stand out. The work is where the machines are: 19 of the 33 said on-site, against 2 remote. Credentials count for more than in AI software generally, with a degree mentioned in 14 and a PhD in 6. And C++ sits almost level with Python (9 against 10 postings), where in AI ads overall Python is far ahead. Python's lower share here partly reflects the non-software roles in the count.
The small print: 33 is a small sample, it leans towards startups and European employers, some Arbeitnow postings are duplicated (so these are postings, not companies), and Australia barely appears. The counts file is public.
What is physical AI, and how is it different from robotics?
Robotics is the older field: machines that sense, decide and act, from factory arms to warehouse carts to drones. "Physical AI" is the newer name for the part driven by learned models: robots that use vision models to see, policies trained by imitation or reinforcement learning to act, and increasingly large multimodal models to follow instructions. In practice the jobs overlap. A physical-AI team still needs classical control and careful safety limits, and a traditional robotics team now uses learned perception almost everywhere.
What skills does a software engineer need to add?
Five, roughly in the order you will need them:
- Simulation. A robot simulator lets you break things for free. Most serious teams test every change in simulation first, and "it works ten times in a row in sim" is a reasonable bar before touching hardware.
- Perception. Cameras, depth, point clouds and SLAM (building a map while locating yourself in it), plus object detection good enough to grasp or avoid something. This is where computer vision skills carry straight over.
- Control. Kinematics, classical controllers and the safety limits that stop a machine from hurting someone or itself. Less fashionable than learning, and the part interviewers probe.
- Learning-based policies. Imitation learning from demonstrations, reinforcement learning in simulation, and the sim-to-real gap: a policy that works perfectly in the simulator and fails on the real floor.
- Deployment on the edge. Running models on a small onboard computer with limited power and memory, and field testing, which means logging failures and explaining them.
Around all of that sits ROS (the Robot Operating System), the common middleware for passing messages between the parts of a robot, and a fair amount of C++ where speed matters.
Do you need a degree or a PhD to work in robotics?
Not always, but more often than in AI software. A degree was mentioned in 14 of the 33 postings (42%) and a PhD in 6 (18%), against 12% and 3% across AI ads generally. PhD requests cluster in research-heavy roles: new learning methods, novel perception, foundation models for robots. Applied roles, integrating perception into a product, building simulation pipelines, deploying and testing on hardware, are where engineers without a research background get in. Only 7 of the 33 postings stated years of experience, with a median of five, so read that figure lightly.
The practical answer: if you have no degree, your portfolio has to carry more weight, and it has to include real hardware, not only simulation.
Can you learn robotics without buying a robot?
You can learn most of it without one, and you should start that way. Simulators, public sensor recordings and open datasets cover perception, control and learning. But plan for hardware before you apply for jobs, because the question every robotics interviewer asks is some version of "what happened when you ran it on the real thing?" A small, inexpensive platform is enough. What matters is the field-test report: what you expected, what happened, what failed and why.
How do you show an employer you can do the job?
With a portfolio that follows a system from simulation to hardware, not a collection of notebooks. The artefacts hiring managers can open: a simulated task that succeeds reliably, a perception pipeline run on recorded sensor data with its accuracy reported, a controller tested under disturbance within safety limits, a learned policy compared honestly against a classical baseline, and the same pipeline on a real machine with a field-test report and a failure log. The comparison against a baseline and the failure log say more about you than a clean demo video.
Where should you start this week?
If your Python is solid, install a robot simulator and get a simulated arm or mobile robot to complete one simple task, then make it succeed ten times in a row. Then add a camera in simulation and detect the object you are reaching for. If your Python is shaky, fix that first; everything above sits on it. Brush up on the linear algebra behind rotations and transforms, because it appears on day one of control and perception.
Where does Square 1 teach this?
The Physical AI and Robotics Bootcamp is twelve weeks, live on Zoom with one instructor, about 15 hours a week, in six blocks that each end in a project and a gate: a simulated robot completing a task ten times in a row, a perception pipeline on recorded sensor data, a controller that stays within safety limits under disturbance, a learned policy that must beat the classical controller, the pipeline on real hardware (the platform is named at enrolment) with a field-test report defended in a recorded viva, and an employer brief with a hiring sprint. It asks for Python and some linear algebra. The Computer Vision Bootcamp builds the perception side in depth, and how its cohort runs has the details. Python for AI is an on-demand course, recorded by an instructor and graded by Nova, the AI tutor, for anyone who needs the Python first. All three are taking a waitlist. The free computer vision skill check takes about three minutes.