--- date: 2026-04-17 --- # 2026-04-17 ### Sleep — fell into the trap again Stayed up until almost 2am building a new robot design. Productive but made today stressful and off. Felt short with Izzy. This is now the 4th late night flagged (Apr 8, 10, 12, 17). The pattern: late-night productivity feels good in the moment but costs more than it gains the next day. The relationship cost is real too. ### Robot arm — design pivot Printed the double-servo design but it's getting complicated. Pivoting back to a **single-servo longer link design** for simplicity. Concern: uncertain whether longer links will be too much torque for current servos (Feetech STS3215). Long-term will likely need to upgrade to Dynamixels. For now: build a shorter single-servo design and print Sunday. ### Panda — taking it back Master's student didn't seem enthusiastic, probably won't take initiative. Reached out to another one, waiting on response. Realistically may just have to do the Panda calibration myself. Plan: take another stab at ArUco-based Panda calibration from scratch. Not that far off — maybe one more day of printing a calibration board, calibrating, and being done. ### Lab meeting + social Good day otherwise. Lab meeting, went out with friends after. ### 3D printer situation Bought a Bambu A1 for lab, but advisor said the lab will just buy one instead. So taking the A1 home and installing it Monday. Means parallel printing at home + lab. ### Weekend plan - **Friday:** Spending night with Izzy - **Saturday:** All-day volleyball tournament + wedding. Will still work between things. - **Sunday:** Back to lab. Order of operations: 1. Print the calibration marker board (small, quick print) 2. While that prints, work on Panda ArUco calibration 3. Launch the big robot arm print 4. Take the A1 printer home 5. Assemble robot arm Monday at home with finished prints ### Evening — Feeling restless but working through it Felt restless today — partly bad sleep, partly feeling stuck. The project feels bottlenecked by robot arm building, which is overwhelming as a first-time build with ~6 weeks to CoRL. But the high-level agent gave reassuring perspective: sim results are strong, we already have some real robot results, the paper stands on its own even without more. More professional-grade real robot results would strengthen it, but it's not do-or-die. ### Robot arm — progressive scaling plan De-risk strategy for the long-arm torque concern: 1. Build a small arm same length as current SO100 (known to work) 2. Build an in-between length arm, see how servos handle it 3. Increase length incrementally 4. If needed, double up servos (already validated the double-servo design) 5. Long-term: switch to Dynamixel motors (already have one at home, same interface across sizes, so switching cost is manageable) For CoRL: the smaller/mid-size arm between SO100 and ARX length should be fine. Don't need the full ARX-scale arm for this paper. ### Research — video + point tracking Recorded hand videos of object manipulations today. Using these to de-risk video generation and point tracking generalization. Good progress toward real robot video model experiments. ### Gym Great 25-min pump session. Fridays aren't busy so got in despite not going early. Quality over duration. ### Reflection — Why do we have multi-day gaps? Two big bottlenecks causing multi-day breaks in progress: 1. **Panda is in the lab** — can only work on it on campus. This is part of the motivation for building our own robots. 2. **Printer access** — lab printer requires scheduling, being on campus, and waiting. Weekend work stalls when you can't print. **Solution:** Now that advisor is buying a lab printer, take the Bambu A1 home. This means: - Print overnight at home (10hr prints run while sleeping) - No need to go to campus just to start a print - Weekend progress doesn't depend on lab access - Can iterate on designs without scheduling around others **Priority:** Get the home printer set up ASAP. This is a force multiplier for everything — every day without it is a day of unnecessary friction. ### Tomorrow (Saturday) — Deep work call agenda Even with volleyball tournament + wedding, get these moving: 1. **Parametric robot URDF** — build a model where you can change link lengths and auto-generate corresponding Blender meshes (link geometries + connector cube geometries). This is the foundation for rapid arm iteration. 2. **Data visualizer tab on server** — new tab/agent on the dashboard for viewing training data. All experiment tracks should push their data visualizations there. Critical for understanding what we're actually training on. 3. **Episode separator tool** — parse iPhone-recorded video into episodes. Set start/end clip segment points in the video for training data. Needed for parsing real-world hand-recorded demos. These three together unlock: point tracking experiments, video model training, and fast robot design iteration. ### Tonight Relaxing with Izzy. Watch TV, have a great night. That's the intention — fun, not work.