Espurna
EN
Learning paths

Session 1: Robotics fundamentals

The full map of a robot: actuators, communication, sensors, mechanics, kinematics and learning. Each piece once, with its demo.

Also available in: Català · Castellano

A robot is always the same loop: sense, think, act. Measure the world (cameras, lidar, encoders, IMU), decide (code, kinematics, learned models) and move (motors, gearboxes, grippers).

By the end you’ll be able to name every piece, explain why a servo is more than a motor, which sensor answers which question, why a leg is not like an arm and how to read a robot as a chain of frames.

Actuators

1
Actuators

How motors work

Current inside a magnetic field = force. Three families depending on what switches the current: brushes (DC), a controller that counts steps (stepper) or electronics that read the rotor (brushless).

2
Sensors

Rotary encoders

An encoder tells the motor where it is. With two phase-shifted channels (A and B) you know how far it has turned and in which direction.

3
Control

Closed-loop control (PID)

A PID compares what you want with what you measure and corrects using three terms: P looks at the present, I at the past and D at the future.

Talking to the motors

4
Communication

How chips talk (UART, I2C, SPI, CAN)

A bus is a set of shared wires. Each protocol has its own way of deciding who talks, to whom and when: UART point to point, I2C by address, SPI with a select wire and CAN by message ID.

5
Communication

Addressing on the servo bus

If six servos share one cable, each needs an ID. They all hear everything and only the matching one replies; two servos with the same ID talk over each other.

Sensors

6
Sensors

Load cell

Strain gauges glued to a piece of metal change their resistance slightly when it bends. A Wheatstone bridge gives a few millivolts, an ADC turns them into counts and calibration converts those to grams.

7
Sensors

Ultrasonic distance sensor

It sends a pulse and measures how long the echo takes: d = c · t / 2. The speed of sound depends on temperature, so if you don't compensate for it the reading drifts.

8
Sensors

Lidar and SLAM

Counting wheel turns accumulates error (drift). A lidar measures distances in many directions at once; SLAM builds the map and locates itself in it by aligning each scan, which corrects the drift.

9
Sensors

Stereo depth

Two cameras set apart see the same object shifted; nearby things shift more. Depth comes from Z = f · B / d.

10
Sensors

IMU and sensor fusion

The accelerometer knows where down is but gets confused when moving; the gyroscope is smooth but drifts. A complementary filter trusts the gyroscope in the short term and the accelerometer in the long term.

Mechanics

11
Mechanics

Gearboxes for joints

A gearbox with ratio N divides speed by N and multiplies torque by N. Planetary, harmonic drive or spur gears: each family trades off precision, robustness and how easily it can be moved from outside.

12
Mechanics

Backlash

There is a little play between gear teeth. When the direction reverses, the motor turns without moving the output, and the error at the tip grows with the length of the arm.

Frames, forward and inverse kinematics

13
Kinematics

Coordinate frames and transforms

A point only has coordinates relative to a frame (origin + axes). To go from the robot's frame to the world's: p_world = R(θ) · p_robot + t.

14
Kinematics

Forward and inverse kinematics

FK: from angles to hand position, always one answer. IK: from position to angles; there can be two solutions (elbow up or down) or none if the point is out of reach.

15
Kinematics

The Jacobian

The Jacobian maps joint velocities to hand velocity (v = J·ω). When its determinant is zero, the arm is at a singularity and loses a direction of motion.

Learning: RL, imitation, VLA

16
Learning

Reinforcement learning

Nobody tells the robot how to do it, only whether it did well: the agent tries actions, gets a reward and learns a policy that maximises it. In simulation, thousands of copies train at once.

17
Learning

Imitation learning

A person teleoperates the robot, observation-action pairs are recorded and a policy is trained to copy them. It fails in states it has never seen; the fix is more varied demos and predicting several actions in a row.

18
Learning

Vision-language-action models (VLA)

A VLA starts from a model that already links images and text and adds an action output: it takes camera images, an instruction and the robot's state, and returns a chunk of movements.