Peer review welcome — the WorkCell series + ACT reproduction report

[ahmedsohail2003]

I'm looking for honest technical feedback on the WorkCell series — five projects teaching one simulated robot work-cell (SO-ARM100 in MuJoCo) with five different approaches — and on the accompanying ACT paper-reproduction report. Everything is public and self-contained; nothing to install.

Start here (5 min): 🌐 ahmedsohail2003.github.io — the whole series on one page, with demos.

Pick your path

🔧 If you build robots (~20–30 min)

• Watch the demo GIFs on the portfolio page, then skim Part A's README (imitation learning + robustness study) and Part E's BUILD_LOG (ROS 2, nine silent-failure modes).
• Questions I'd love answers to: Where would this break on a real arm that the sim is hiding? Is the domain-randomization range (mass ×0.5–2, friction ×0.7–1.4, servo gain ×0.8–1.2) meaningful or toy? Does the asymmetric-gripper TCP story in Part A's README match your hardware experience? What's the first thing you'd test that I didn't?

🎓 If you're academically inclined (~30 min)

• Read the ACT reproduction report — a claim-by-claim reproduction of Zhao et al., RSS 2023, against my own experiments.
• Questions: Is the verdict table's REPRODUCED / PARTIALLY / NOT TESTED scoping honest? Are the statistics handled defensibly (n=20–50, single seed, Wilson CIs — §1 and §7)? Is the temporal-ensembling sign-flip extension (§5.2) a real finding or over-interpreted? Would you accept §6.3's "coverage whack-a-mole" as a negative result, and what control is missing?

💻 If you're a software person (~15 min)

• Skim any repo's README top-to-bottom, then one script (e.g. eval_policy.py).
• Questions: Can you follow what was built and why from the README alone? Where did you stop reading, and what lost you? Anything in the code you'd flag in review?

How to leave feedback

• Reply to this thread (any length — bullet points welcome), or
• Open an issue on the specific repo if it's a concrete bug, or
• Email me if you'd rather not post publicly: ahmedsohaill2003@gmail.com (personal — preferred) or butt4320@mylaurier.ca (school)

Blunt criticism is the useful kind — "this claim is too strong", "this would never survive contact with hardware", "section X reads as filler" all help me more than praise. Thanks!