Refraction AI

I started working in Refraction as an intern and later transitioned into a full time employee. I met some amazing people, learned a lot and built some cool robots.

Firstly Refraction AI builds autonomous delivery robots. Unlike smaller sidewalk robots our robots travel on the road margin (bike lanes/ side of the road), which means we can be bigger and move faster.
Being in a startup and given my skillset, I worked on everything from building hardware to writing autonomy software. Below is a brief writeup of some of the stuff that Ive done (obviously I took permission before posting this stuff) :)

On starting my full time job at refraction, one of my first few projects was to build a whole new hardware platform from scratch. We were a bunch of software engineers but were up for the challenge :) Luckily a few of us knew CAD while the others learnt PCB designing and electrical design.

I focused on building the base platform while my colleague worked on building the sensor stack (which enables autonomy). We went from prototyping with cardboard, to building a test bot with a shower curtain to finally actually molding the panels of the bot and assembling the whole thing together. This spanned around 6-8 months.

Initial Chassis Testing, I built up the rig to hold the sensor suite

Initial duct tape body, I built up the internals with pvc to hold it up

Testing model, made out of fabric (we used a shower curtain! cause it was waterproof)

I tried and tested out multiple body configurations, along with chassis configurations and door configurations before actually reaching to a conclusion for out final body design. Below are a few renders of what I tried out.

Initial render of body designs

Sensor FOV depiction

The main subsystems I worked on the new platform was-
1. The brake subsystem
2. The steering subsystem

I designed them in house and got them manufactured. Both worked extremely well. I learned about bleeding brakes, found the perfect tolerance and assembly tactics :)
One of the other things I did as well, was help with assembly for the whole fleet of robots :) We unfortunately did not have the budget to get it outsourced and since I was part of the design team I volunteered to assemble the bots. This gave me amazing insight into what to change for the next iteration to make assembly quicker and faster. But yea with help from my teammates we had 12 bots up and running :)

Me literally in the bot :)

Testing out wire routing + electronics and controllers!

All our robots, from first generation (in the left) to third generation (in the right)

Robot testing

You have to keep in mind that we were a software focused company, so after deploying our 3rd generation platform I was back to writing autonomy software. I wrote some embedded code to add some features to the inbuilt keypad on the robot. It uses CAN and helped with robot proprietary functionality. Apart from this I worked on tuning the controllers for the robot so it drives as intended. I also worked on shaving down docker image sizes for streamlined deployment and accelerated iteration with minimal overhead.

Other software projects that I worked on include an intuitive visualization tool for sensor extrinsic calibrations, allowing for quick identification of any inaccuracies. I also implemented a streamlined pipeline for efficiently accepting and rejecting calibrations on the robot, enhancing our overall workflow and ensuring more reliable performance.

Those were the necessary building blocks to reach one step closer to autonomy. With those in place I worked on actual autonomy. The core of my work can be showcased in the two videos below

UKF based lane tracking and reference line generation

Behind the scenes

The videos above essentially show the robot detecting lane lines using a neural network, projecting it to 3d and then tracking those lines based on a UKF. If you have more questions on this feel free to contact me :)

Well I did a lot more in refraction that I could not document here. We used a super small budget and built some pretty amazing things, and I will always be grateful for working there!!