Aditya Kothuri

About

Aditya Kothuri

Robotics & ML engineer

placeholder

Robots during the week; mountains whenever the forecast allows. This page is the part of the story that does not fit in a results table.

Aditya Kothuri standing in front of a partly frozen alpine lake, with a pine forest and a snow-covered mountain ridge behind.
Aditya Kothuri

Home was a place where the horizon mattered — flat in every direction, which is probably why elevation became the hobby. The first things I built were not robots; they were excuses to take something apart and see whether it went back together better.

Robotics arrived sideways: a simulation that refused to match the machine it claimed to model, and the slow realisation that the gap between the two was more interesting than either side of it. That gap is still where I work.

Off the clock I walk uphill. Long routes, cold lakes, the kind of terrain I now spend weekdays reconstructing — which means a trip is never entirely off the clock, because every ridgeline is also a heightfield I have opinions about.

placeholder — personal copy to be replaced

How I work

One pipeline, five stages. Each consumes what the last produced, which is why the order matters and why a failure at the second stage stays invisible until the fifth.

  1. 01

    Capture

    Real-world footage and public elevation data, reconciled against each other where they disagree — and they always disagree.

    Photogrammetry · DEM rasters

  2. 02

    Reconstruct

    Point cloud to mesh to heightfield, resampled with a filter that preserves sharp features instead of rounding them away.

    Open3D · NumPy · GDAL

  3. 03

    Simulate

    Terrain imported as a collider a solver can actually step over, at a cell size matched to the robot's contact patch.

    Isaac Sim · USD · PhysX

  4. 04

    Train

    Reinforcement learning with randomization wide enough that the policy survives the gap — and verified by reading the spread back out of the simulator, because a config that claims to randomize is not the same as one that does.

    Isaac Lab · PyTorch

  5. 05

    Validate

    Judged by something that shares no code with the thing being judged: a second physics engine, or a held-out benchmark with real ground truth. Failed runs stay in the log with their takeaway.

    MuJoCo · ONNX

Experience

  • 2025 — presentplaceholder

    Robotics & ML engineer

    Independent

    Simulation and reinforcement learning for legged locomotion on reconstructed real-world terrain, and inference-time research on vision-language models.

Off the clock

  • Alpine routes

    The long way up, ideally with a cold lake at the top.

    placeholder

  • The workbench

    Small machines, taken apart on purpose and reassembled on principle.

    placeholder

  • Photography

    Mostly terrain. The camera goes where the boots go.

    placeholder

Contact