The dock
The arm docks the drone.
The drone lands on a pad within the arm’s reach, and the arm lifts it by its handle into a cradle and back. The arm moves only while the drone reports disarmed, and the drone arms only while the arm rests.
Robotics & machine learning
Sim to real.
In two volumes.
One method, two machines. In each volume you write the simulator, test classical controllers, train a model that adapts to physics it has never seen, and run it on hardware you designed, printed and built. Volume 1 builds a desk robot arm. Volume 2 builds a drone.
Start on your laptop. Build the hardware when you’re ready.
FIG. 001 / ADAPTIVE ROBOT ARMDESIGN STUDY
02volumes: arm, then drone
18build modules
25written episode companions
01method, two machines
The volumes
Start with the arm.
Then fly the drone.
Two volumes, one question: what must a network learn to adapt to physics it has never seen? Volume 1 is where to start: its kit costs about €505 in parts including a 3D printer, and it needs a desk, not a flight space. Volume 2 builds on the same method and on the arm itself.
Both volumes
Arrow keys move between the volumes.
Volume 19 modules · 12 episode companions
Adaptive Robot Arm
Write an arm simulator, train a model that adapts, then design, print and build a desk robot arm and run the model on it.
Open Volume 1Volume 29 modules · 13 episode companions
Adaptive Quadrocopter
Build an adaptive quadrocopter from first principles: write the simulator, train the controller, then print and fly the aircraft.
Open Volume 2How the volumes connect
One desk.
Two machines.
Volume 1 fixes the interface both volumes design against, so the arm you build stays on the desk when the course moves on to the drone.
The test stand
The torque stand becomes the thrust stand.
The load cell, HX711 and bench Pico 2 that identify the arm’s servos in Module 8 measure the drone’s motors in Volume 2.
The method
The second volume starts from what you know.
Both volumes write their own simulator, keep one benchmark from Module 2 on, and end every module with a measurable gate. Volume 2 was written first; its spec gets the docking interface when it is revised.
One question runs through both volumes
What should a network learn
to understand a robot’s hidden physics?
Payloads change. Servos differ, and their firmware is closed.
Pick a training objective, then build the experiment in Module 4.
Before you begin
A few good
questions.
Do I need a robot arm to start?
No. Modules 1–5 run on your laptop. You write the simulator, test classical controllers and train the adaptation model before buying hardware. You design the arm yourself in Module 6, which can start after Module 1.
Which volume should I start with?
Volume 1, the arm. Its kit costs less, it needs a desk rather than a flight space, and Volume 2 builds on it: the arm becomes the drone’s docking station, and the torque stand you build for its servos becomes the drone’s thrust stand. Both volumes follow the same nine steps, so the second one starts from methods you already know.
How much experience do I need?
You should be comfortable with Python, matrices, and basic derivatives. The course introduces control theory, reinforcement learning, CAD, and the hardware build as you need them.
What will it cost?
The written course is free with an account, and the software uses open-source or free tools. Volume 1’s kit costs about €505 in parts including a 3D printer, or about €315 if you already own one with a 180 mm bed. Volume 2’s drone kit costs about €1,160–€1,200. Both are planning estimates before shipping and tools, not quotes.
Do I need a GPU?
A laptop CPU is the intended path. The training budget is under two hours, with a compact model of at most 10,000 parameters. You will measure your actual runtime and simplify the model if needed. Volume 2 adds an optional comparison with a rented GPU as a separate experiment.
What is in the starter kits?
Each volume has one, free with an account: the benchmark harness, the gate checks and the progress log. Volume 1’s adds the task checks and the safety supervisor every powered move goes through. You write the simulator, controllers, estimators, learning code, CAD and firmware yourself.
Are the episodes and customer tiers available?
The written module guides and episode companions of both volumes are here. The specs also map out free recordings, customer recordings and trainer materials. Recorded videos and managed customer access are not available yet, and nothing here needs a checkout or subscription.
What counts as completing the course?
Each module ends with an evidence-based gate. You record a reproducible run, submit the measurements to the gate check, and document the result. Hardware modules also need logs and a video for review. Marking a page as studied tracks your reading; it does not certify a gate.
Volume 1 starts on your laptop