Posts
-
How do you know your robot got better? Part 2: measuring the evaluation
Companion to Part 1 — training a diffusion policy on a $300 arm from 0% to ~43%. This part is about the “~”. We set out to rank six model variants and ended up measuring our evaluation instead — at four levels, down to the physics of why two identical grasp attempts disagree 27% of the time. If Part 1 was about making the robot better, Part 2 is about how hard it is to know whether you did.
-
Teaching a $300 arm to pick up a block, part 1: training
For people running robot-learning experiments on cheap hardware (SO-101-class arms) who want to train a simple model themselves rather than fine-tune a giant VLA. This is a journey post: the dead ends are the content, not the final recipe. Part 1 covers getting the policy from 0% to ~43%. Part 2 covers what “~43%” even means — how we ended up measuring the evaluation itself, down to the physics of why two identical grasp attempts disagree 27% of the time.
subscribe via RSS