Month 6 of 6
Robot learning, specialisation, and becoming hireable
Robot learning, specialisation, and becoming hireable
You now have the whole stack: electronics, embedded, mechanical, ROS 2, control and perception
This month has two halves. The first is the frontier, which is teaching robots from demonstrations rather than programming them. The second is turning everything you have built into something that gets you hired
Part one: modern robot learning
This is the part of robotics that changed completely in the last two years, and it is the reason there is so much capital in the field
1LeRobot
LeRobot is Hugging Face's PyTorch library for real-world robotics, and it is the open hub the whole low-cost robot-learning world has converged on
The workflow is: teleoperate, record, train, deploy. You drive the robot by hand, demonstrations are saved as synchronised video and action data, a policy learns to imitate them, then it runs on its own
Resources
LeRobot documentation (free)
The main docs, covering the full pipeline and every supported robot
LeRobot repository (free, Apache 2.0)
Over 26,000 stars, and the place to read how the policies are actually implemented
Hugging Face Robotics Course (free, no hardware required)
Runs entirely on simulated environments and public datasets, so you can do the whole thing before buying anything. Roughly 30 to 45 minutes per unit
SO-101 setup guide (free)
The exact port-finding, motor-setup, calibration and recording commands for the arm you built in month three
What to focus on
Practice task
record 50 demonstrations of a single simple task on your SO-101, such as picking up a cube and dropping it in a bin, train an ACT policy, and deploy it. It will work maybe half the time. Then record 50 more demonstrations covering the failure cases and retrain. Document the success rate before and after. That number, and the fact that you measured it, is the portfolio piece
2Vision-language-action models
These are the foundation models of robotics, and knowing which ones you can actually run matters
- π₀, π₀-FAST and π₀.₅ from Physical Intelligence, Apache 2.0, weights open, pretrained on 10,000+ hours of robot data. The repo warns honestly that these were developed for their own robots and transfer is not guaranteed
- OpenVLA, 7B parameters, fully open, trained on 970,000 robot episodes from Open X-Embodiment, and the best-documented open VLA to read the code of
- GR00T N1.7 from NVIDIA, code Apache 2.0 with weights under the NVIDIA Open Model License, which is a distinction often reported wrongly. Needs 16GB+ VRAM for inference
- SmolVLA from Hugging Face, compact and designed for affordable hardware, which is the right one to fine-tune on an SO-101 rather than reaching for a 7B model
- RT-2 from Google DeepMind is historically important and has no public weights, so study the paper and use one of the above for practice
3Reinforcement learning for robotics
Resources
MuJoCo Playground (free, Apache 2.0)
GPU-accelerated environments for locomotion, manipulation and vision tasks with four Colab tutorials, and far easier to get running than the alternatives. Start here
NVIDIA Isaac Lab (free, BSD-3)
Sixteen robot models and thirty-plus pre-built training environments, integrating RSL RL, skrl, RL Games and Stable Baselines. The industry standard for legged and humanoid sim-to-real, and it needs the RTX hardware from month four
CS 285, Deep Reinforcement Learning, Sergey Levine, UC Berkeley (free)
The best RL course available, and Levine is a robotics researcher so the framing is robotics-native throughout, covering imitation learning, policy gradients, actor-critic, model-based and offline RL
Practice task
train a quadruped locomotion policy in MuJoCo Playground from one of the Colab tutorials, then change the reward function and observe how the gait changes. You do not need hardware and you do not need a GPU beyond what Colab gives you free
Part two: pick a direction
Three directions genuinely exist. Pick one, and let the other two stay at literacy level
Direction 1: Robot learning and embodied AI
Best if you want the frontier companies and the highest ceiling
Focus on: LeRobot, VLA fine-tuning, imitation learning, RL, simulation, PyTorch
This is the highest-paid track and also the most competitive, and it is where a real dataset you collected yourself is worth more than any credential
Direction 2: Autonomy and mobile robotics
Best if you want the largest number of available jobs
Focus on: ROS 2, Nav2, SLAM, perception, sensor fusion, C++
Controls engineer and field service engineer together are over 20% of all robotics postings, and this direction serves both
Direction 3: Embedded, mechatronics and integration
Best if you want to work immediately, including as a contractor
Focus on: firmware, motor control, real-time systems, PLCs, functional safety, system integration
This is the least glamorous and the most consistently employable, and it is where the technician-to-engineer path actually runs. Note that 66% of robotics projects report delays caused by certification, so functional safety knowledge is a genuine and underserved specialism
4Your portfolio
I looked at what recruiters in this field actually say they screen for, and the pattern is consistent
High signal
- GitHub repos with logged metrics and a commit history that shows iterative debugging, not a finished demo dropped in one commit
- Evidence of real-robot deployment with reliability data, not simulation only
- Public datasets on the LeRobot Hub, where there are already tens of thousands of community datasets
- Contributions to the packages employers depend on: ROS 2 core, Nav2, MoveIt, Isaac Lab, LeRobot
- Systems integration work joining sensors to actuators to planning to control, with video and clear documentation
Red flags they name explicitly
- Still on end-of-life ROS 1 with no migration evidence
- Claimed projects that cannot survive three follow-up questions
- A pure deep-learning background with no kinematics or embodiment understanding
- Tutorial completion certificates, which are weighted far below real contributions
Practice task
take your three best projects and rewrite their READMEs. Each one needs a video at the top, a wiring or architecture diagram, the actual numbers you measured, and a section titled "what broke and how I fixed it". That last section is the single highest-value thing in a self-taught portfolio, because it is the part that cannot be faked from a tutorial
5Interviews
Robotics interviews are not software interviews, and leetcode is a much weaker predictor here
Expect: inverse kinematics questions, PID and feedback loops, sensor fusion and Kalman filters, SLAM concepts, path planning including RRT, and C++ versus Python trade-offs
Expect also, at good companies: a simulation-based debugging exercise where they hand you a MuJoCo or Isaac scene with a deliberately broken controller, a ROS 2 architecture problem scoped to the real job, and a long conversation about a specific past failure and how you diagnosed it
Glassdoor has 1,721 robotics engineer interview questions on file from 877 companies if you want to read real examples
glassdoor.comPractice task
have someone interrogate you about your own repo for twenty minutes. Not the concepts, your specific code. Why that gain, why that sensor, what happens if the battery sags, what did you try before this worked. If you cannot answer three levels deep, the project is not ready to be on your CV
Month 6 Milestone
By the end of this month you should be able to:
Check yourself
2–3 hours a day is the design load.