We ran TACT-ful on the actual robot for the first time a few weeks ago, stairs with 20 cm risers, the humanoid carrying a 20 kg box, zero retraining from sim. I was standing by the stairwell with three other people from the team, everyone pretty quiet, and it just went up. No fine-tuning, no last-minute checkpoint swap. One try, and for a second nobody said anything because we half expected it to trip on the last step.

I've sat through enough of these demos by now that I probably shouldn't get this worked up over a policy transferring cleanly, but I still do. Most of what changed for this one was in the terrain cost maps and how we simulate carrying a load without actually modeling contact forces. None of it feels especially clever looking back. It's more that we finally stopped guessing at what mattered and started testing it properly.

Was cleaning out an old drive and found the code from my undergrad thesis: a structured-light setup for picking parts out of a bin with an industrial arm. Looking at it now it's kind of rough, hardcoded thresholds everywhere, a calibration routine I had to babysit by hand every time the lighting changed. But I remember being genuinely thrilled the first time it picked the right part on its own.

It's a strange thing to compare to what I work on now, since RL for humanoid locomotion doesn't look anything like structured-light scanning on paper. Most days though it still comes down to a similar question: does the robot know enough about what's around it to act correctly. Back then that meant getting a clean point cloud out of two cameras and some gray codes. Now it means training a policy across a few thousand simulated environments and hoping it holds up outside of them. Different tools, but the same feeling when something finally works that didn't the day before.