Robotics papers decoded, machines priced, and the chip-industry job market counted — translated from our Korean field lab.
The five ideas that unlock every robotics paper of 2026 — world models, VLAs, simulation, sim2real, and the data problem — explained with no math and linked to real decoded papers.
KAIST’s HOUND took the Science Robotics cover by fusing walking, running, jumping, and ledge-climbing into a single Transformer controller — with 15.5 hours of training data generated in 8 minutes.
TouchWorld imagines the pressure it is about to feel and treats that feeling as the goal — beating Pi-0.5 and GR00T N1.7 on contact-heavy tasks, and degrading least when humans interfere.
A Science Robotics paper from Zhejiang University shows a low-cost robot hand learning bottle caps, levers, and pencil sharpeners from one webcam and a touching/not-touching signal — about 73% overall success.
One training run, one neural network — and 12 differently shaped humanoids walk, including robots it never saw. A plain-English decode of the cross-humanoid whole-body control preprint XHugWBC.
World-action models could imagine the future but were too slow to walk with. MotionWAM grabs the video model’s half-finished ‘sketch’ in a single pass and hits 4.9 Hz on a real Unitree G1 — 76.1% average success across 9 tasks.
800 hours of first-person human video plus just 30 hours of robot data beat rivals trained on ten times more. A plain-English decode of the RSS 2026 humanoid foundation model paper Ψ0.
A Nature Machine Intelligence paper ran 600+ experiments to find which VLA design choices actually drive robot performance — backbone quality, continuous actions, and history fusion. Decoded in plain English.