Simulation, Learning and Edge AI at Arrive AI

What I’m doing

I joined Arrive AI after finishing my PhD, and the work is a direct continuation of it: simulation first, hardware second, with the gap between them as the actual engineering problem.

Simulation at scale. Building a digital twin environment in NVIDIA Isaac Sim so scenarios can be tested before anything reaches the real world. This is the part of my research I most wanted to keep doing. A simulator earns its keep when it catches a failure that would otherwise have happened in front of someone.

Reinforcement learning. Continuing the policy-training work my PhD was built on.

Vision-language-action models. Standing up and evaluating open VLM and VLA models such as OpenVLA and OpenPI.

Perception and inference. Evaluating and optimizing vision transformers for real-time perception, accelerating inference with NVIDIA Triton, the DeepStream SDK and TensorRT.

Choosing the silicon. Profiling edge devices for the latency-throughput trade-offs that decide what will actually run: Jetson Orin Nano, Orin AGX and Thor (Blackwell), and, beyond NVIDIA, Qualcomm’s IQ9075 and hardware from Infineon Technologies. Picking a compute platform is a decision you live with for years, so it’s worth measuring rather than assuming.

Getting it deployed. Architecting the containerized AI inference codebase with Docker so deployment is repeatable rather than artisanal, working with the platform team on the k3s side. I also build the web frontend for low-latency WebRTC video streaming.

About Arrive AI

Arrive AI connects people, robots and drones through the Arrive Point™ Network, built to scale the future of delivery. This page describes my work at the level it appears on my resume. The systems themselves aren’t mine to show, so there’s no footage or product detail here. If you’d like to talk about any of it in more depth, get in touch.