During my three month internship, I ran a total of 89 experiments (most with multiple iterations over some selection of hyperparameters), had 261 commits across 7 branches of the project codebase, and gained valuable experience on the know-how for an interdisciplinary machine learning research project in industry.
Working in the software team of Project SARAM, I:
- delved deep into RL literature and theory (self-studied all of this specialization) and state of the art from OpenAI (etc. etc. etc.) which we based our work off on
- fine-tuned existing models, reducing average episode length for the task by ~25% while investigating potential approaches for sim2real transfer
- incorporated the newest URDF robot model and object meshes for the env
- developed the new simulation env, adding robust env dynamics/agent configurations and maintaining good code modularity
- proved that the robot is able to learn solely via exploration, prioritize and plan its own trajectories while producing human-like stirring techniques
Here’s the presentation deck I gave at the intern showcase to over a hundred people including Samsung Research America’s President, Joon Lee.
P.S. I’m not sure if anything in the deck violates NDA. I’ll leave it up as I don’t think any approaches alluded to is patented, and the RL work was based off of OpenAI’s public papers. However, if anyone from the team chases me down, I’ll take it off immediately.