day 40: from zero to humanoid went dark for 40 days. spent all of it going deep into humanoid robotics and understanding how all of this actually works now and how it needs to work for much efficient outputs. here's what i figured out. (some key takeaways, not all) • the bottleneck isn't ai. it's actuators. everyone's obsessed with the brain but torque density is the real war. quasi-direct drive gives you back-drivability and impact tolerance, harmonic drives give you precision but explode on falls. every company is quietly making a different bet here and it defines their whole robot. • hands are 50% of the problem. a humanoid body is around 30 degrees of freedom. one human hand is 27. tendon-driven hands wear out in weeks, linkage hands can't do fine manipulation. nobody has solved dexterity + durability + cost at the same time. whoever does wins general-purpose robotics. • sim2real is a leaky abstraction. you can train locomotion in isaac gym in hours, but contact dynamics, cable friction, and motor thermals don't transfer. domain randomization papers over it. the gap between "walks in sim" and "walks in your kitchen" is where 90% of engineering time goes. • the data problem is worse than anyone admits. llms had the internet. humanoids have nothing. teleoperation gets you demos at ~$X/hour of human time, and you need millions of hours. this is why everyone's fighting over VLA architectures and they're trying to steal generalisation from vision-language models because robot data doesn't exist. • thermals are a silent killer. a humanoid standing still is burning power just fighting gravity. knee and hip motors saturate thermally in minutes under load. this is why demo videos are short. battery is 1-2 hours real-world, and adding battery adds mass which needs more torque which needs more battery. • whole-body control is having an identity crisis. classical MPC is precise but brittle. end-to-end RL is robust but unexplainable. the frontier is hybrid and RL for locomotion, model-based for manipulation and nobody agrees where the seam should be. the takeaway: this industry needs shared knowledge infrastructure more than it needs another robot company tbh that's what i'm building. more soon.