Machine learning does exactly what you tell it to

This list of published machine learning fails is pretty entertaining. There are a few classes of errors:

• agents exploiting bugs in physics simulations;
• underspecified objective function results in a useless result;
• in one terrifying case, overcoming a physical limitation: "A robot arm with a purposely disabled gripper found a way to hit the box in a way that would force the gripper open"

Project log: Smart robot car (part 2)

Over the years I've ordered many dev kits. So far I've only used my Arduinos and Raspberry Pis.

But today is the day: I'm firing up an ESP8266 mini.

This ended up being non-trivial.

• A Beginner's Guide to the ESP8266 by Peter P: too long, wanted quick instructions to flash my Wemos D1 Mini with a blink sketch.
• Wemos Wiki: pointed to standard Arduino board support package on github.
• ESP8266 Arduino Github repo: decided to start here. The Arduino IDE is a pain to use, but I wanted to just get to the point of flashing the D1. This did not work for me, and the Arduino IDE took seconds just to open menus! Luckily there was a pointer to a makefile...
• github.com/plerup/makeEspArduino: this one seems to require the IDE was set up. Before I tried it, I decided to Google for an alternative.
• Compiling for the ESP8266 with Arduino and OTA support without the Arduino IDE sounded promising. One that was actively maintained, and the author's own fork that was "easier to use." Actively maintained please.
• github.com/thunderace/Esp8266-Arduino-Makefile 

End result: no success while trying to avoid curl | bash. I'll have to resort to that.

Project log: Smart robot car (part 1)

I decided it would be fun to do some computer vision + robotics. Why not a smart car?

The first step was to get a car chassis that could be controlled by microcontroller. A quick stop at Amazon and Adafruit got me a chassis complete with motors and wheels and a motor driver.

Assembling the chassis was straightforward, but the next step was controlling the motors using a microcontroller. I'd want wireless control eventually, but first I wanted to work with a familiar MCU. So I pulled out an Elegoo Arduino Nano and wired things up with the help of Adafruit's tutorial:

I had a dual motor controller that could provide ~1A/channel; four wheels/motors, each drawing up to 1A; and a 12V battery pack that could only supply 2A. I wired the wheels on each side up in serial. That should result in the two motors on each side splitting power, so the driver isn't overloaded.

I started off stripping individual wires to size, but when I got to the actual signal wires I decided it was going to be a little messy regardless, and it was temporary anyway. Once the 12V power supply and motor wires were in (had to add in some male-to-female jumpers for those because the wires I added were too short!) it really was a mess.

It was about this time that I decided I should put in an order for a Wemos D1 motor shield, so I'd have the option for simpler wiring later on. It'll take 2-3 weeks to get here, but I've got the Adafruit breakout in the meantime.

With a bit of code stolen from The Blaster I was manually flipping bits to control motor direction in no time. To simplify things a bit for the future, I added a new header with a TB6612 class. This class provides forward(), backward(), short_brake(), and stop() methods that flip the appropriate bits. No PWM support for now.

Next, switching to an ESP8266 or ESP32 for WiFi, so the car can be controlled remotely.