Throw like you Catch
For our final project in 6.843 = Robotic Manipulation by Russ Tedrake - I wanted to make robots play catch. By the means of Inverse Kinematics, optimization based closed loop DiffIK, we planned an interception strategy to catch a Tennisball from various locations. Try to catch a Ball with chopsticks. Its hard. I found that playing the videos backward made it look like throwing, so given a target position and velocity, we can reverse the velocity and solve the catching problem. Reversing our plan will yield in a throwing controller! Run this Deepnote project to find out for yourself! Or read the details.
This summer, I bought a motorcycle on craigslist in California. The Goal: Make a cafe racer and drive it back to Boston. Using this bike for my daily work commute got me stranded a couple of times until I decided to take it apart and reverse engineer it. Instead of fixing the electronics I redid everything and threw the ignition lock away. It has now Keyless go, the blinkers are well hidden, minimalistic buttons control everything, and check out the speedometer! I broke my ankle in a skydiving accident, which made me transform my dorm room into a workshop until I got the last fixes done.
16.001 - Flight Competition
Our final project was to design and build a plane that maximizes the payload carried and the angular velocity is a circle squared. I was responsible for the optimization of the wing. We ran a simple linear optimizer for each plane configuration and since some trends were obvious we were able to reduce the hyperparameter search space to less than 700.000 planes. It took my server 3 days to find the optimal solution :)
Later we build the plane and I made sure that all the wings are smooth
2.S007 - Challenge
The class taught us how to design and build a Robot to fulfill specific requirements. My robot has a four-bar structure to couple the scooping mechanism with the slide such that the sand can be transported to the gripper. Further, I enabled some autonomous tasks like pulling the rocket, scooping and grabbing objects
16.07 - PID control
One of our Labs for the intro to Controls was to balance a planar helicopter with a control system. This video demonstrates PID with feed forward control.
Pencil Picker ;)
RRT for motion planning
1. Goal: Grab a pencil and put it upright
2. Use Computer Vision do find bounding box of the object
3. Find the transformation to the goal
4. sample face on bounding box and find valid initial and final grasp poses
5. Use RRT to find a path between initial and final pose
The first design project for my 6.08 class required us to develop a simple IoT enabled watch. But I overkilled it and invented a whole new concept of time in a weekend.
Nexttime has two states:
Now: displays current time, date and current calendar event with progress
Next: A minimal countdown to the next calendar event including relevant info.
With a simple server, an esp32, and a well organized calendar the countdown to the next event is the only time that I keep track off.
FEA tools for 3D Frames
I developed a Space Frame analysis tool in python. After taking a class in structural engineering, I wanted to build an electric elevating bed. Instead of doing the analysis by hand, I decided to develop a finite element analysis tool using the direct stiffness method to optimize materials, beam properties and further parameters of my bed. I build the entire bed within 24 hours before my flight to Boston. This was only possible through a very detailed CAD. GitHub
Fraunhofer IPK (2020)
My task was to improve the local OCR methods s.t. an independence from cloud vision solutions is gained. The existing implementation used a text localization network which feeds in the snipped of the image to an extractor (pyTesseract).
Since pyTesseract text extraction is very volatile to distorted images, I figured that combining many detectors yielded better results than google cloud vision.
I implemented the following detectors:
NEET Virtual Competition
Fall 2020 was online for all MIT students. Our Neet seminar send us kits to build and program autonomous robots. Besides making a bot follow a line, we had to sense colors on a board and execute a sequence of commands as fast as possible. My robot won the color challange.
Staging Demonstrator 2
This year our Rocket Team traveled to Lancaster in California to test our Staging Demonstrator in preparation for Project Spaceshot. We successfully proved that our ignition system works. That was crucial for starting the design process on a much larger two-stage Rocket.
Enjoy the video!
FIDL - Face ID Door Lock
2019 - 2020
After moving to MIT I kept losing my room keys over again. So why not using Facial Recognition to unlock the door?
My Roommate was Immediately fascinated by the idea of opening and closing our door with only pull and push motions.
The main components are a raspberry pi 4, the google edge TPU, a camera, a display, a servo motor, and some 3d printed parts.
I designed a multiple-stage AI, which is first detecting faces in the camera image, classifying each face, and deciding wheater Access is Granted or not. The most difficult part was to develop a fine-tuned multiprocessing architecture to classify our faces in real-time, opening the door immediately, and hosting a server for diverse web applications.
MIT Maker Portfolio
2 minutes for an application video sound short, but I could handle it. So I decided to rap it as quickly as I can!
Lab coat dispenser
2 in 1 couch table entertainment system
MINI Server hosting a VPN, Nextcloud, Blynk, etc.
IoT Gate Opener (GPS, App, Voice assistant, and Jedi triggered)
Active noise cancelation for Windows
Custom surface transducing speakers
App-enabled Robot arm (feeding me with cereals )
Wireless payload dropping mechanism playing happy birthday (gift delivery)
Modified Ninja claws
Soldering Tin smoke blaster
Research on theft secure bicycle
Remote Arduino Compiler
Blindspot assistant for motorcycles
Recycled BBQ fire igniters
MIT Fun Video
In addition to my Portfolio, I submitted a longer video about my first Rocket!
With restricted resources and regulations, I was still able to derive a solid propellant. This video will also show you the Rocket I build and my custom Avionics System for verifying real and estimated flight trajectories.
With a 1090 MHz trimmed Antenna, a software-defined radio (RTL-SDR), and a raspberry pi I was able to decode transponder information. The data I gathered is in real-time on my Server in my hometown Berlin.