Undergraduate Robotics Researcher · Mar 2025 - May 2025

IDEAS Lab

Built a real time SLAM pipeline in C++ and Python with RGB D sensor fusion and Kalman filtering, streaming calibrated camera data to a multi GPU server, improving 3D reconstruction accuracy 25% and reducing latency 30%.

  • Implemented neural radiance fields (NeRF) for novel view synthesis, generating photorealistic 3D scene reconstructions from camera data for autonomous navigation.
  • Generated real time navigation signals from incomplete 3D maps, choosing where a robot should explore next in environments it has not fully mapped.

What's actually hard about ML research

It was the model, to be honest. Trying to write a model with minimal public code out there, outdated libraries, and duct taped GitHubs that were written in languages people don't use and set up for specific GPU jobs was a pain. There needs to be a solution to the problem of being able to easily reproduce and extend ML models, beyond just downloading the weights.

Reproducibility in robotics

Robotics research has the worst version of this problem. The methods for getting a trained model running on a physical robot are barely documented, and training pipelines are rarely reproducible without rebuilding the whole rig from memory. What's different about robotics is the hardware coupling and sensor calibration. Setting up RGB D cameras is not easy, and neither is real time streaming the data to a server across multiple GPUs.

What I'm proudest of

Generating real time navigation signals from incomplete 3D maps. There's so much that goes into figuring out where to go when you're exploring. I didn't even get to work on it for long enough to produce results that I was happy with, but identifying how to optimize navigation and spatial search in unknown environments is a genuinely applicable problem.

What research lab work is really like

It's really hard. It's so hard being thrown into the deep end of having to understand papers, find code, stitch it together, try to run tests, and iterate on difficult to set up models. Having a good GPA doesn't help with that, which is what many people get wrong. You need a genuinely creative mindset, and so much grit and persistence, enough to actually start making progress, because starting out basically feels impossible.

Slowly, then all at once

The all at once moment came late. We made nearly all our progress in the last three weeks. Progress on a new project is always like this: slowly, then all at once. Once you've tried enough things, pieces just start clicking together, and progress comes quickly once the bottlenecks clear.