Xingxin on Bug Scientific Method
Coming from a design background, I was never formally trained in the scientific method. Perhaps the concept was introduced passively in some general courses, but the first time I officially learned it was from the book Microeconomics by Daron Acemoglu. This week, I finally managed to close a full loop: observation, hypothesis testing, and experimentation.
Seeing my hypothesis proven correct was a truly inspiring moment for me!
Scaling Law
The biggest lesson learned this week is that I should try multiple policies right at the beginning. I didn’t truly grasp the concept of “scaling law” taught in Stanford CS336 | Language Modeling from Scratch. My initial plan was to use ACT as the starter policy to find the optimal data mixture ratio before scaling up to NVIDIA GR00T and OpenPI.
But I missed one crucial thing: when things went wrong, I didn’t try switching to a 📄Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. It wasn’t until this week that I realized Diffusion Policy yields much better performance for my setup than ACT. The cost of learning this lesson was two weeks of time.
Gripper Issue
From the last 2 weeks, I have been having trouble with the gripper values in both the observation space and action space. Specifically, a tiny calibration error or tolerance would shift the predicted values far away from the standard deviation, which causes even worse issue in ACT than Diffusion Policy.
Luckily, my colleague suggested I take a look at an OpenPI Github issue that discusses a similar problem. This is amazing work from the open source community!
Windsurfing
This Saturday marks the first time I used a 4.2 sail!! Goodbye 3.5 :- )
I also learned some easier rigging steps from my coach. I usually struggle with tying the sail to the tail of the boom, but after this session, I think I finally nailed it!
