1x speed, not sped up. The arms line up flower stems and thread them into a narrow clear glass vase on a lab table, depth-view inset running alongside. The kind of final-millimeter placement that usually gets cut from demo reels.
The model behind this is LingBot-VLA 2.0. pi-0.5 and GR00T are open too, so open isn't the difference; what Robbyant reports as different is 20 robot configurations through one 55-dim action vector, plus whole-body control extended to head, waist, mobile base, and dexterous hands, and predictive-dynamics pretraining.
Robbyant notes it often makes partial progress but fails at the final precise placement or release. That limitation is why watching it attempt the vase insertion matters more than a clean success reel. You see the near-miss surface in real time.
So is this real generalization, or another benchmark sculpted to look autonomous until the final millimeter? Their self-reported GM-100 eval has one long-horizon task dropping from 60.0% to 13.3% success out-of-distribution. Does that gap mean anything, or is it the same benchmark-shaping with more hardware in the pool?