For those that don’t want to touch the GPIO header of their Jetson
Made with @BackplaneWorks
STM32G474
3x CAN-FD
3x LSM6DSV16X IMU
+ legacy IO
My $1,065 PCBA run got hit with $229.21 of taxes and fees at the border: 21.5%
Even after reclaiming GST, that's 11.5% in fees on goods with a 0% duty rate.
Why? I was $65 over the import threshold. It was $1,000 in 2005. It's $1,000 today.
I was born in 2005.
JLC X-ray inspection is pretty cool
The LGAs meant to be inspected are fully shadowed by the LQFP on the opposite side but I guess that’s ok?
Jack Wilson retweeted
explained my agent setup at dinner and my family looked at me like I joined a cult
How the test bench started, how it’s going now.
The first time Claude flashed my STM32 I was in awe
Now it tests a bus of my smart sensors while I sleep
Sometimes it needs me to do a power cycle
I’m just the meat that presses the button
Time to get a programmable power supply
ALT Prototype sensor nodes, STM32 Nucelos, all over CAN bus to the dev server
Testing some ROV thrusters from Taobao.
1/8th the price of a Blue Robotics T500. Corrosion resistance and reliability are yet to be seen, but they definitely push water.
Follow along as I build an autonomous underwater robot with them.
Can an underwater robot learn to see from renders alone?
Testing on real images: First model, 0.22 AP50. A week and 100k more training images, 0.94.
This is what made the transfer stick
3. Just let your agent rent 15 4090’s
- Decided to scale up. Initial dataset was 10k images
- With formula paying off, tried scaling up dataset size
- Rendered in IsaacSim, agents make scene layouts and rent runpods
- 106,000 images rendered on 15 GPUs in 1 hour, cost $30
4. Thoughts
After scale up, reached 0.94 on real dataset
Main takeaway for me - how fast the feedback loop was. I’d have a hypothesis, an agent would generate new data and train - I would have a result in 45 minutes.
Datasets at huggingface.co/jack-www
