Posts

I used three foundation models – SAM, Grounded DINO, and FoundationPose – to obtain 6D pose estimates on raw RGBD robot demonstrations. In order to improve results, I added L1-consistent bounding boxes and various annotations to the data. Ultimately, the results were mixed, and we’re still looking for improvements.

When I first started studying reinforcement learning (RL), I implemented Proximal Policy Optimization (PPO) from scratch using only the psuedocode on OpenAI’s website. It didn’t work and failed to obtain nearly any reward on most OpenAI Gym environments.

The Google Foobar challenge is a programming challenge used by Google for recruiting. You must be invited to participate (article with more info). The challenge contains five levels of increasing difficulty, which require knowlege in the following areas: