cv

Master of Science

University of Colorado Boulder

• Computer Science

• Robotics and ML systems. Research in the Boulder AI and Systems Infrastructure Lab (BASIL) with Prof. Mark Zhao.

Bachelor of Science

The University of Applied Sciences Bingen, Germany

• Computer Science

Bachelor of Engineering

SRM University, Chennai, India

• Computer Science and Engineering

Research Intern

AMD Research and Advanced Development

• Sim-to-real RL for the XArm-6 with MuJoCo Warp on Instinct and Radeon GPUs, deployed on the real arm with a classical perception front-end.
• Co-authored the AMD ROCm blog on training and deploying the pick-and-lift policy.
• Upstreamed ROCm and RDNA4 fixes to PyTorch (CK SDPA flash attention on gfx1200/gfx1201) and AMD Research's Ryzers (Drake on ROCm, GPU-accelerated robotics stacks on Strix Point).

Graduate Researcher

Human Interaction and Robotics Group, CU Boulder

• Contact-aware controllers and safety filters for human-robot collaboration (CERG).
• Prototyped algorithms in simulation (Drake, NVIDIA Isaac Sim) to help robots differentiate between objects and humans.

Graduate Researcher

Boulder AI and Systems Infrastructure Lab (BASIL), CU Boulder

• Systems for deploying transformer models on edge NPUs (SHARD, EuroMLSys'26).
• GPU-parallel robot simulation and sim-to-real RL with MuJoCo Warp.

Embedded Software Developer

Trinamix GmbH (BASF), Ludwigshafen, Germany

• Led the adoption of a new Rockchip platform, cutting prototyping costs by 90% through hands-on board bring-up, device tree configuration, and low-level hardware debugging (I2C, GPIO).
• Developed Linux V4L2 camera drivers in C/C++ to integrate custom laser sensors, enabling computer vision-based liveness detection for a major German automotive OEM.

Intern, Bachelor Thesis

NXP Semiconductors, Hamburg, Germany

• Implemented novel calibration optimizations for an NXP automotive radar front end, developing C-based test firmware.
• Achieved >50% reduction in calibration time through algorithmic improvements.

Unlocking Sim2Real for a Robotic Arm with RL Accelerated by AMD Instinct GPUs

AMD ROCm Blogs (co-author)

Arkojit Ghosh, Adithya Mohan, Noah Monti, Mehdi Saeedi, Sarunas Kalade, Graham Schelle

Compliant Explicit Reference Governor for Contact Friendly Robotic Manipulators

IFAC World Congress 2026, Busan

Yaashia Gautam, Gilberto Briscoe-Martinez, Adhitya Mohan, Nataliya Nechyporenko, Alessandro Roncone, Marco M. Nicotra

SHARD: A Compatibility Framework for Deploying Transformer Models on Edge NPUs

EuroMLSys'26 (Sixth European Workshop on Machine Learning and Systems at EuroSys)

Adhitya Mohan, Richard Thompson, Eric Keller, Mark Zhao

LIBERO on MuJoCo Warp - 130 manipulation tasks on one GPU

• Ported all 130 LIBERO manipulation tasks to MuJoCo Warp (MJX + Warp), batched across thousands of parallel worlds.
• 17,137 physics env-steps/s at 2,048 worlds on a $700 RX 9070 XT, against 38 env-steps/s on CPU.
• Fixed the Warp ray tracer to match MuJoCo's CPU renderer so a CPU-trained BC transformer transfers (42.5% vs 50% on CPU). Writeup: https://amohan.dev/blog/2026/libero-warp-mjx-rdna4/

SHARD: End-to-End VLM Inference on Rockchip RK3588

• Reverse-engineered undocumented NPU hardware limits (32KB SRAM, 4KB Page) to deploy a VLM that failed with standard vendor tools, creating novel 'Nano-Tiling' and 'Graph Surgery' techniques.
• Achieved <2.0s latency (a 15x speedup over CPU baseline) while maintaining FP32-equivalent fidelity (>0.999 cosine similarity), enabling real-time deployment on commodity hardware.

OpenNPU - an open compiler and runtime for the RK3588 NPU

• An open compiler and runtime that runs transformer models on the RK3588 NPU without the vendor toolchain.
• Runs GPT-2 at 36 tok/s on the NPU. Writeup: https://amohan.dev/blog/2026/opennpu-v1-0-open-compiler-runtime-rk3588-npu/

FP8-as-Storage GEMM for Ampere (RTX 3090)

• IMMA-based FP8-as-storage GEMM experiments for Ampere (sm_86 / RTX 3090 Ti), no H100 required.
• Writeup: https://amohan.dev/blog/2026/fp8-as-storage-imma-ampere/

Neuro-Symbolic VLM for Adaptive Impedance Control

• Designed a system connecting a VLM's scene understanding to a low-level impedance controller for a Franka Emika arm, validated and tested in NVIDIA Isaac Sim.
• Boosted material classification accuracy from 28% (pure visual) to 100% by developing a neuro-symbolic pipeline that combines OpenCV geometric detection with VLM reasoning via scene graphs.

Compliant Explicit Reference Governor (CERG) for Contact-Rich Tasks

• Designed a safety filter that enables robots to operate safely during physical contact by placing a hard limit on the total system energy, ensuring provably safe interaction.
• Published as arXiv:2504.09188.