Skip to content

Aakash Madabhushi

Robot Learning

I train manipulation policies on a physical arm and build the evaluation apparatus that says honestly whether they work.

Graduate Research Assistant in the SJSU Department of Applied Data Science since March 2026, training and evaluating SO-101 pick-and-place policies on real hardware. The work covers the whole loop: teleoperated data collection, quality gates over the recorded episodes, PyTorch and CUDA training runs, and a physical evaluation harness that scores rollouts one at a time and survives being interrupted.

Training the policy is the easy half. The harder half is the measurement: a harness that returns the arm to start on its own, guards against stale caches, logs every failure mode separately, and reports 92% instead of rounding it up. That work now extends into Isaac Sim and MuJoCo for SO-ARM101 and Franka Emika Panda arms, so control approaches can be checked before they reach hardware.

Email meRésuméAll projects
92%
pick-and-place success, 50 physical rollouts at the trained cube position
2.49 cm
average placement error on successful trials
49,633
demonstration frames published as an open dataset, across 50 episodes

3 projects in this area

Every link goes to code or data you can inspect yourself.

The SO-101 arm on the lab desk beside the marked cube it picks up and the clear container it places the cube into.
Robot LearningML Engineering

Teaching a Robot Arm to Pick and Place

Trained a real robot arm to pick up a cube and place it from 50 human demonstrations, then built a resumable evaluation harness and scored 50 physical rollouts at 92% success.

92%
pick-and-place success, 50 rollouts at the trained cube position
2.49 cm
average placement error on successful trials
50
teleoperated episodes published as an open dataset, 49,633 frames
  • Python
  • PyTorch
  • CUDA
  • LeRobot
  • ACT
  • SmolVLA
  • Weights & Biases
  • Hugging Face Hub

Graduate Research Assistant, SJSU Applied Data Science · Mar 2026 to present

Read the case study View code on GitHubACT policy on Hugging FaceTraining data, v2Training data, v1
The physical robot arm on the lab desk beside the monitor running the simulator it is matched against.
Robot LearningSoftware Engineering

Real2Sim2Real: Mirroring a Real Workspace into Simulation

Locates the real cube and bowl with printed markers and a calibrated overhead camera, then places them at matching coordinates in a physics simulation, so policies train against the real table's layout.

  • Python
  • NVIDIA Isaac Sim
  • ROS 2
  • OpenCV
  • ArUco
  • +2 more

Personal project, built on NVIDIA's Sim-to-Real SO-101 workshop · Aug 2026 to present

View code on GitHub
Robot LearningSoftware Engineering

ROS 2 + Gazebo, Containerized

Runs ROS 2 Jazzy, Gazebo Harmonic, RViz, and a TurtleBot3 model from one command, identically on macOS and Linux. A built-in desktop makes the GUI tools work without fighting X11.

  • Docker
  • Docker Compose
  • ROS 2 Jazzy
  • Gazebo Harmonic
  • TurtleBot3
  • +1 more

Personal project · Mar 2026

View code on GitHub

Tools I've shipped with here

Policy learning

  • PyTorch
  • CUDA
  • LeRobot
  • ACT
  • SmolVLA
  • Imitation learning
  • Teleoperation

Simulation & robotics

  • NVIDIA Isaac Sim
  • MuJoCo
  • ROS 2
  • Gazebo
  • OpenCV
  • ArUco

Experiment tooling

  • Weights & Biases
  • Hugging Face Hub
  • Docker

Relevant experience

The parts of my roles that bear on this work. The résumé has all of it.

Graduate Research Assistant · SJSU Department of Applied Data Science

Mar 2026 to present

  • Built an end-to-end LeRobot v3 data pipeline for SO-101 pick-and-place imitation learning, synchronizing 6-DoF joint state and action trajectories with dual 640x480 RGB streams at 30 FPS, and published 50 teleoperated episodes (49,633 frames) to Hugging Face.
  • Developed pre-training quality gates over Parquet metadata and H.264 video, validating episode and frame alignment, video decodability and resolution, and NaN-free action and proprioceptive tensors.
  • Engineered a reproducible PyTorch/CUDA training pipeline for 52M-parameter Action Chunking Transformer policies with configurable augmentation, Weights & Biases tracking, 5K-step checkpointing, and resume support, completing 30K- and 60K-step runs.
  • Built a resumable physical-robot evaluation pipeline for a 100-trial fixed and randomized cube-placement protocol with automatic return-to-start, stale-cache guards, placement-error measurement, and success/grasp/failure logging. Completed all 50 fixed-position rollouts at 92% success and 2.49 cm mean placement error.
  • Building physics-based manipulation simulations in Isaac Sim and MuJoCo for SO-ARM101 and Franka Emika Panda arms, validating control algorithms before deployment.

Hiring for robot learning?

I'm available from December 2026 and open to relocating. Email me and I'll reply within a day.

aakash.vardhan15@gmail.com

Hiring for something else?

Data EngineeringAI EngineeringML EngineeringSoftware Engineering