
5/5 Before blaming the sim-to-real gap, read the simulator. Full post, with @wenhaocha1 and @andylizf: p-b-p-b.github.io/blog/2026/…

CS PhD @ Princeton

5/5 Before blaming the sim-to-real gap, read the simulator. Full post, with @wenhaocha1 and @andylizf: p-b-p-b.github.io/blog/2026/…

4/ Values no real thing has: five of the arm’s six joints may turn 573° each way. The real arm stops within half a turn. About half the benchmark’s released demonstrations bend a joint further than the real arm can. The garbage is already in the dataset.

3/ Checks that pass without doing the task: the mallet passes over every key without touching one, and the task credits all eight strikes. Shoving bottles off the table into the bin, picking none up, earns up to 40 points.

2/ One task gives its bottles a friction of 10,000, stickier than glue. We asked the physics engine what it was actually using: 0.5! The benchmark attaches materials after the engine has loaded the objects, so the engine never sees them, for any object in any task.

What does RoboDojo really simulate? We followed it from the real world to its score, one link at a time. At almost every step, it simulates something other than what it says. Same bottle, same grasp. As shipped, it tips 78° in the hand. After our fix, it stays level.
