Aman Arora
Software Engineer
Hire this person
Send a job offer directly to this candidate
About
Software Engineer building reproducible reinforcement learning systems and environments for real-robot control. Developed GPU-parallel training pipelines in Isaac Gym and MuJoCo MJX, implemented PPO, SAC, and constrained variants from scratch, and diagnosed training failures including synchronization bugs. Four years of production systems engineering at NVIDIA and Philips.
Experience
Senior Machine Learning Engineer - Syntiant - Bay Area
(2026-06)
Designing ML and RL Training and Deployment Pipelines for power efficient peripheral models and devices.
Research Assistant - Quadruped Robot Learning / Sim2Real - University at Buffalo, TRAIL Lab
(2024-08 - 2026-06)
- Built an end-to-end quadrupedal locomotion reinforcement learning system spanning simulation, policy training and evaluation, failure debugging, and sim-to-real deployment on Unitree GO2 hardware.
- Designed a GPU-accelerated reinforcement learning environment in Isaac Gym for large-scale parallel simulation and policy evaluation.
- Implemented teacher-student locomotion learning using imitation learning (DAgger) with domain randomization for sim-to-real transfer, enabling policies trained with privileged simulation signals to generalize to observation-based real-robot control.
- Built data-collection, evaluation, and experiment-tracking workflows that improved reproducibility and failure analysis across robotics model iterations.
- Achieved locomotion performance including stair climbing up to 23 cm at 38° incline and obstacle traversal up to 30 cm in simulation, with student policies transferring to real-robot deployment.
- Diagnosed and fixed an actuator-network training failure by correcting a data-collection thread synchronization bug, reducing MAE from 3.0 to 0.15 and restoring convergence.
- Mentored new lab members on simulation environment setup, training infrastructure, and reproducible robotics experiment workflows.
System Software Engineer - GPU Compiler (PTX / PTXAS, CUDA Toolkit) - NVIDIA - Bangalore, India
(2019-03 - 2023-07)
- Designed and maintained core components of the PTXAS GPU JIT compiler within the CUDA toolkit, delivering production-grade compiler functionality across GPU generations.
- Implemented compiler support for Hopper warp-group matrix multiply-accumulate (HGMM(A)) instructions by defining instruction semantics and refactoring instruction handling for large combinatorial shape variants through scalable parsing and validation.
- Co-inventor on 3 US patents for Hopper tensor-core operations, including asynchronous warp-group pipelines enabling memory-compute overlap.
Software Engineer - Philips Ultrasound R&D (DICOM Component Owner) - Philips Innovation Campus - Bangalore, India
(2017-07 - 2019-03)
- Developed components for a clinician-facing ultrasound system built from scratch using a shared-memory IPC message bus architecture; resolved a system architecture issue by redesigning message-bus lifecycle initialisation, eliminating repeated bus creation during user interactions.
Education
M.S. - Computer Science and Engineering - Research Track - University at Buffalo (2023-08 - 2025-06)
B.Tech - Computer Science and Engineering - Maulana Azad National Institute of Technology (MANIT) (2013 - 2017)
