Who it is for
Robotics integrators and automation teams without large in-house demonstration-collection programs.
Why now
Two of the three constraints moved in this window while the third was measured and stayed put. Streaming tactile conditioning now conditions action generation on evolving tactile observations so a separate high-frequency controller is unnecessary, and a pixel-motion action interface plus an emergent inverse mode claims transfer of human object flow to robot actions without task-specific demonstrations. Against that, a new dexterous visual-tool-use benchmark found 25 evaluated models bottlenecking at 0.40-0.54 reconstruction similarity with closed-loop instability, and a curriculum-trained 9B model improved only 6.3% relative — so the honest reading is partial relief, not collapse.
What would kill it
Evidence that sim-to-real transfer fails on contact-rich tasks or that streaming tactile and pixel-motion interfaces do not generalize beyond their evaluated benchmarks, which would leave the demonstration-collection cost intact.