Day 3: I wanted a plan and got a shopping list instead

Another day, more time spent in the conference room than anywhere near a workbench. Turns out that’s just how this stage of the project goes — more questions than answers, and today’s questions were the expensive kind.

We started with the brain: what does the compute unit actually need to do, how much headroom does that take, and — the part that matters — how long before I outgrow it. Honestly, I didn’t expect this hardware to run into the hundreds of dollars. As a hobby project, I’d mentally budgeted maybe $20-50 a month. Today’s spec put the number at $600, from the very start, and something tells me that’s not the final total.

Raspberry Pi — the obvious starter board everyone reaches for first — got ruled out almost immediately: not enough headroom for where this is going. That left two real contenders from Nvidia’s Jetson Orin line: the NX and the Nano. And here’s the part I didn’t see coming: buying the Nano is just the entry ticket. It’s enough for one full module — one arm’s worth of joints, or one leg’s. Only one. One arm. Or one leg. Uff. Good news, if you can call it that: none of these boards become dead weight later. Scaling up just means buying more of them, coordinated by one NX sitting above the rest. Far future. We’ll see.

So, the Jetson Orin Nano it is, for the next several stages. Bonus discovery: there’s a “Super” Dev Kit version that already bundles the carrier board you’d otherwise have to source separately — apparently that’s the entire difference between the dev kit and just the module. Glad I found that out before ordering the wrong one.

Then came the camera, and I drowned. Depth of field, focal length, resolution, measurement accuracy, capture modes and their trade-offs — thank god I have a professional team on this. Short version: I need this thing to actually measure depth, not guess at it. Depth comes in three flavors — approximate (a single camera guessing), computed (two cameras doing math), and measured (a sensor with real per-pixel depth built in). Approximate and computed both land around 2-4cm of accuracy. For something that’s supposed to move like a human arm, that’s nowhere near good enough. Imagine trying to pick up a glass of water, or spear a piece of melon with a fork, with 4cm of precision on your own hand. Death from starvation, or fainting from dehydration, are both realistic outcomes. I’d like better for the robot.

So I went looking at the RealSense D435i, the depth camera everyone recommends, and discovered it’s quietly gotten 2.5x more expensive over the last couple of years — now priced at around $450, roughly what I’m paying for the entire brain of this robot. For a camera. I kept looking, found two other contenders — the Orbbec Gemini 336 and the Luxonis OAK-D Pro — and picked the Gemini: it measures real depth in hardware instead of computing it, throws in an IR filter, and comes in over $100 cheaper than the RealSense.

Goal achieved, more or less. Compute and camera, locked in. Which counts as a good day — if you squint past the fact that I set out to build a development plan and walked away with a shopping list instead.