Progression

Robbyant's LingBot-VLA 2.0 demo shows one trained policy driving everything from a Franka single arm up to Fourier GR-2 and Unitree G1 humanoids with dexterous hands. The clip is all marked 1x speed and fully autonomous. Training mix is about 60k hours, 50k real robot across those 20 configs and 10k egocentric human video. The honest numbers are what make this worth watching though. Generalist success on the Agilex bimanual setup sits around 34 percent, drops to about 15 percent on Galaxea R1 Pro, and some tasks flatline at zero. The authors themselves note it often gets most of the way through a task then fumbles the final precise placement or release. That gap between looking capable and actually finishing the job feels like the real story for VLAs right now.