16-825 Learning for 3D Vision · Fall 2026
I asked Claude to cast the Wason selection task as a robot failure-analysis problem.
My prompt: "Can you make a Wason test in a robot failure-analysis context?"
Claude: The rule is "fall → pitch spike," so the only two cards that can produce a counterexample are the antecedent-true card (a fall — does the IMU back it up?) and the consequent-false card (no spike — did that run somehow still fall?). Everything else is logically inert. That failure mode has a real analogue in postmortem work: if your hypothesis is "the falls come from ankle-torque saturation," the instinct is to pull the runs that fell, but those can't kill the hypothesis. The run that saturated and didn't fall carries the information — and nobody queries it, because it's filed under "successful run."
Full conversation: shared chat
The Klein bottle: a closed, non-orientable surface with no inside or outside and no boundary. It cannot be embedded in 3D without self-intersecting, so every physical "Klein bottle" is really an immersion — a 4D object squeezed into three dimensions.
"Klein bottle" by Tttrung, via Wikimedia Commons, licensed CC BY-SA 3.0.
Boston Dynamics' Atlas attempts a box vault and reunites, face-first, with the crash mat.
Atlas robot, footage © Boston Dynamics, via Tenor.