Advanced Arduino: Dual AS5048A Magnetic Encoders via SPI

This video demonstrates how to use Direct Memory Access (DMA) to implement multiple SPI ports on a PIC32 microcontroller for driving LED strips. The technique involves using a timer to generate DMA transfers that output data to parallel ports, with a secondary DMA task stopping the timer when complete. This approach eliminates the interrupt overhead that limits traditional interrupt-driven methods to around 100 kHz, enabling operation at 500 kHz SBI clock rates. The solution uses a single chip with no external gating, making it simple, low-cost, and easy to debug.

This video demonstrates how to implement non-blocking SPI communication using interrupts on STM32 microcontrollers, where an external interrupt (from a button press) triggers an SPI read operation to fetch device status from an MCP2515 CAN controller module. The implementation uses HAL_SPI_TransmitReceive_IT() for concurrent data transmission and reception, with completion callbacks handling CS pin management and data processing, enabling efficient, responsive embedded systems that don't block the main program execution while waiting for peripheral responses.

Using DMA (Direct Memory Access) for SPI communication with an ST7789 display on STM32 microcontroller significantly improves performance by enabling parallel processing, where the microcontroller core continues executing code while DMA handles data transfer, resulting in approximately 47 frames per second compared to 43 fps without DMA, and maintaining 32 fps even under CPU load versus 31 fps without DMA.

SPI (Serial Peripheral Interface) is a synchronous serial communication protocol developed by Motorola that enables full-duplex data transfer between a master controller and multiple peripheral devices using four signal lines: SCLK (clock), MOSI (master out slave in), MISO (master in slave out), and CS (chip select). The protocol operates in four modes determined by clock polarity (CPOL) and phase (CPHA) parameters, with Mode 0 (CPOL=0, CPHA=0) being the most commonly used configuration where the clock idles low and data is sampled on the rising edge. In this tutorial, we demonstrate SPI implementation on an STM32 Nucleo board using STM32CubeIDE and the HAL library, including both blocking and interrupt-driven approaches for reading from and writing to an external EEPROM chip, showing how to configure the SPI peripheral, manage chip select lines, and implement non-blocking communication for efficient real-time applications.

DMA (Direct Memory Access) allows the ESP32 to transfer large amounts of data directly between memory and peripherals without involving the main CPU. This feature, typically found in more powerful processors, enables smooth animations on TFT displays, audio playback through I2S, and high-speed sensor operation without blocking the main code. With DMA handling heavy data transfer tasks, the main cores remain almost completely free for other operations.

Encoder mounting errors (constant angular bias in joint angle measurements) have the largest effect on positioning accuracy. A two-stage calibration strategy addresses this: First, estimate only the encoder bias parameters (Delta theta) while assuming nominal DH parameters are correct. This removes the dominant error source. Second, restart the full calibration problem with the corrected encoder biases, allowing variations in all parameters. This conditioning improves numerical stability because all parameter variations now have comparable expected amplitudes.

The Denavit-Hartenberg (DH) representation is a systematic method for describing the position and orientation of robotic joints using four parameters: a_i-1 (link length, distance between joint axes along the common perpendicular), alpha_i-1 (link twist, angle between joint axes about the common perpendicular), d_i (link offset, distance between origins along the previous joint axis), and theta_i (joint angle, rotation about the current joint axis). For prismatic joints, d_i is non-zero while theta_i is zero; for revolute joints, theta_i is non-zero while d_i is zero. These four parameters are organized into a DH table, which enables the calculation of transformation matrices that describe the spatial relationship between consecutive frames in a robotic manipulator.

After determining theta1, theta2, and theta3 from wrist center position, perform forward kinematics to compute transformation matrix 0T3 (from base to wrist center). This matrix contains rotation part 0R3. The rotation matrix from wrist center to end effector (3R6) is found by multiplying 0R6 (known from orientation input) by the transpose of 0R3 (since 0R3 × 3R6 = 0R6). This provides the orientation relationship needed for solving remaining joint angles.

Angular position in radians is calculated by scaling the timer counter value using the encoder resolution. Angular velocity is obtained by computing the difference between consecutive angle measurements. Special handling is required during rotation transitions to handle counter wrap-around correctly. Real-time monitoring involves initializing the encoder, reading the timer counter periodically, calculating position and velocity, and optionally plotting results. This feedback system enables precise control applications like inverted pendulum stabilization.

This segment demonstrates applying forward kinematics to an RPR (Revolute-Prismatic-Revolute) manipulator. The DH parameters table provides values for each link-joint pair. For each transformation matrix Tᵢ₋₁ᵢ, substitute the corresponding DH parameters into the general equation. When alpha or a equals zero, those transformations become identity matrices and can be omitted. The final composite transformation matrix T₀₄ contains three variables (θ₁, d₂, θ₃) representing the three degrees of freedom. Substituting specific joint values allows calculation of the end effector's position and orientation in space.

This segment covers the complete process of designing a digital low-pass filter to reduce noise in MPU6050 sensor data. The video demonstrates how raw sensor readings become unstable when the device is shaken, and explains the mathematical foundation using the transfer function G(s) = 1/(τs + 1), where τ = 1/(2πfc). For a 1 Hz cutoff frequency, τ ≈ 0.159 seconds. The instructor shows how to implement this in MATLAB using tf() and c2d() functions, converting the continuous-time filter to discrete-time for digital implementation. This provides a complete theoretical and practical foundation for understanding digital filtering in embedded systems.

Flight controllers receive noisy sensor readings from gyroscope chips due to electrical interference, motor vibrations, and inherent sensor imperfections. Directly using raw measurements causes motors to react to useless garbage data, potentially burning out motors. The process of smoothing noisy data and estimating actual values is called filtering. A PT1 filter balances between trusting new measurements (Dig) and trusting previous filtered values (Bill). With 50/50 trust levels, the filter outputs the average of new and old values. The formula is: new_filter_value = previous_filtered_value × Bill_trust + new_measurement × Dig_trust. Higher Dig coefficients mean more trust in new values (lower latency but more noise), while lower Dig coefficients mean more trust in conservative estimates (smoother signal but higher delay).

A Kalman filter is an optimal estimator that combines predictions from a system model with sensor measurements to produce more accurate state estimates than either source alone, using a two-step process where the prediction step forecasts future states and calculates uncertainty, while the update step combines predictions with measurements based on their relative uncertainties using a Kalman gain matrix; unlike static low-pass filters, Kalman filters dynamically adjust their filtering characteristics based on the relationship between prediction accuracy and measurement reliability, making them particularly effective for RC quadcopters where they can filter sensor noise while preserving dynamic response during maneuvers.

This comprehensive section covers the complete fundamentals of sensor filtering for robot control in FRC. It begins with the importance of sensor filtering for precise robot control, introduces various sensor types (encoders, potentiometers, cameras, IMUs), and explains how optical encoders work by counting transitions on a spinning disc. The section addresses sensor imperfections, including the inability to determine exact position between transitions and the calculation of velocity from position changes, which produces jagged, noisy data. It covers approaches to handling sensor data: ignoring data for low-precision applications, buying better sensors, building stiffer mechanisms, making mechanisms self-centering, improving wiring, and using filtering to remove high-frequency noise. The section emphasizes using prior knowledge about motors, mass, and gear ratios to improve estimates. It then transitions to the Kalman filter, explaining Gaussian noise characteristics, covariance matrices, and the filter matrices (F, H, Q, R). The section details the prediction and measurement update steps, explaining how the Kalman gain determines how much to trust the sensor versus the model based on current uncertainty and sensor noise characteristics.

This tutorial covers two fundamental digital filtering techniques for removing noise from accelerometer sensor data. The Exponential Moving Average (EMA) filter applies recursive weighting using an Alpha coefficient (0-1), where higher Alpha values reduce filtering strength. The Simple Moving Average (SMA) filter calculates the arithmetic mean of a fixed number of recent readings (window size or taps). Both filters attenuate high-frequency noise while preserving low-frequency signal components. The choice between filter types depends on the application requirements: EMA offers computational efficiency with adjustable filtering strength, while SMA provides straightforward averaging with predictable frequency response characteristics.