Arwun, and why I'm putting it on Isaac

This is the first entry in a build log for Arwun — a small, self-balancing tracked robot with flipper-gripper arms — and a place to think out loud while I move its perception stack onto the NVIDIA Isaac platform.

Where the robot is right now

Controls are prototyped in Drake with an LQR design, and I’m wrapping up tabletop bench validation with real servos and an IMU. The plan from here:

• keep Drake as the low-level controls and simulation brain,
• add Isaac ROS perception GEMs on the Jetson Orin Nano as measurement sources,
• feed those into PRISM, my factor-graph world model.

Why Isaac, and why now

A few reasons, in order of how much they actually drove the decision:

1. The split is convenient. Isaac ROS GEMs lean on the GPU; PRISM’s GTSAM optimization sits on the CPU. On an 8 GB Orin Nano that division of labor matters.
2. It’s the canonical sim-to-real loop. Simulate in Isaac Sim, generate synthetic data, deploy to the edge via Isaac ROS — and I want to measure the gap honestly rather than eyeball it.
3. It speaks the language I’m aiming at. No sense being coy: this work doubles as a portfolio for physical-AI roles, so building on the actual stack is the point.

What I expect to go wrong

If past Jetson work is any guide: version drift between CUDA, TensorRT, and ROS, and the usual fight to expose the GPU inside Docker. I’ll write those up as I hit them — that’s the part future-me will want.

Next entry: standing up the Isaac ROS container and running a single GEM against a rosbag.