Industry Experience

Nextdoor

Learning how to train and serve ML models at scale.
Machine Learning Engineer: July 2023 - Current
Machine Learning Intern: Summer 2022

8th Wall (Now Niantic)

Software engineer working on AR products: Summer 2021
Software engineer working on fullstack development: Summer 2020

Projects

AlphaGarden - A Turing Test for Automated Gardening

Applying machine learning and computer vision to automate polyculture gardening.
  • Simulation The AlphaGardenSim is a fast, first order simulator, that incorporates parameterized individual plant growth models, companion plant effects and inter-plant dynamics, and was leveraged to learn irrigation and pruning policies using deep RL.
  • Real to Sim Computer vision models and algorithms to interface between the physical garden and simulation.
  • Sim to Real Custom hardware and action planning algorithms are used to enact both learned and analytical policies from AlphaGardenSim into the real world.
  • Real World Experiments We present 8, 60 day garden cycles, where we compare different pruning and irrigation policies versus expert human horticulturalists.
As a kid, I remember making my mom drive me to the library every month to read the next copy of popular science magazine, So you can imagine my excitement that a project I worked on made it into it's own article!
PopSci article

Learning Switching Criteria for Sim2Real Transfer

Learning in the real world is expensive. It takes a long time to run real world robot experiments, and has potential to break parts or cause damage.

More commonly we use Sim2Real transfer to learn in simulation and deploy the same model into the real world using domain random- ization, behavior cloning or other strategies. In this project, the team explored stopping criteria for when to learn in simulation and when to switch over to the real world for optimal performance.

Here, I worked with Softgym and GymCloth - simulators made to train RL agents on tasks such as cloth folding. I worked to shape the codebase to our needs and created various fidelity environments to learn in. I also built out utility functions for helping perform RL in Softgym.

To evaluate in real we worked with a physical robot setup, where I helped debug and code a test suite using a YuMi robot.

CS294 – Geometry and Learning for 3D Vision

As an undergraduate, I took a grad level course on 3D computer vision which explores recovering three-dimensional scene structure and camera motion from multiple two-dimensional images. Some concepts covered include epipolar geometry, camera calibration, SLAM, and more.

I applied the skills I learned in this class to explore object tracking for applications in autonomous vehicles in a group of 3 students. We explored using scene flow for rigid body object tracking from point clouds without annotations and used warm start techniques to increase optimization speed. The final algorithm achieved comparable AMOTA (accuracy) and AMOTP (precision) for single class object tracking on the Nu-Scenes benchmark to state of the art learned models.


CS 294 - Augmented / Virtual Reality

I took a course in Augmented / Virtual Reality where I was able to extend my knowledge of computer vision and computer graphics in a multi-discipline class mixing computer science and design. As a part of my final project, I worked with two design other students to explore applications of stable diffusion and virtual reality. We explored how img2img, text2img models and audio recognition can be used in VR. Leveraging the Oculus and open sourced painting application - tiltbrush and image generative models.
This demo was presented at an end of semester showcase, where users were able to play around with the product.


Computer Graphics

Last school year I took a computer graphics class where I explored rendering and graphics using C++. Here I worked on several projects including a raytracer, first order cloth simulator, and more. In this class, I learned a lot about cool rendering and graphics techniques, as well as the geometry behind it. At the end, I used what I learned to implement a 3D soccer simulator with graphics in C++ with 3 other classmates.
Project Links:
Soccer Sim
Cloth Sim
Path Tracer
Mesh Editor
Rasterizer

Robotics

Designed, built, and programmed a robot!

Automated block pickup and driving to aid in human navigation across the field. Developed coarse and fine grained controllers using encoders, cameras, and ultrasonic sensors.
Won NASA's Engineering Inspiration Award.

(See our robot in action Team 2854)

Publications

  • Simeon Adebola*, Rishi Parikh*, Mark Presten, Satvik Sharma, Shrey Aeron, Ananth Rao, Sandeep Mukherjee, Tomson Qu, Christina Wistrom, Eugen Solowjow, Ken Goldberg
    "Systematically Comparing Growth and Irrigation Performance: The AlphaGarden vs. Professional Horticulturalists" in Conference on Robotics and Automation (ICRA)
    * = Equal Contribution

  • Mark Presten Rishi Parikh, Shrey Aeron, Sandeep Mukherjee, Simeon Adebola, Satvik Sharma, Mark Theis, Walter Teitelbaum, and Ken Goldberg
    "Automated Pruning of Polyculture Plants; Mark Presten" in 2022 International Conference on Automation Science and Engineering (CASE)


  • Satvik Sharma, Ellen Novoseller, Vainavi Viswanath, Zaynah Javed, Rishi Parikh, Ryan Hoque, Ashwin Balakrishna, Daniel Brown, Ken Goldberg
    Learning Switching Criteria for Sim2Real Transfer in 2022 International Conference on Automation Science and Engineering (CASE)

  • Y. Avigal, A. Deza, W. Wong, S. Oehme, M. Presten, M. Theis, J. Chui, P. Shao, H. Huang, A. Kotani, S. Sharma, R. Parikh, M. Luo, S. Mukherjee, S. Carpin, J. H. Viers, S. Vougioukas, and K. Goldberg
    “Learning seed placements and automation policies for polyculture farming with companion plants,” in 2021 IEEE International Confer- ence on Robotics and Automation (ICRA), 2021, pp. 902–908.