ROBOTS
The robot that beat Usain Bolt still can't plug in a cable
Tiangong ran 100m in 8.86s at Beijing's 2026 robot games — then robots caught fire. Why stop distance and manual override matter more than sprint records on site.
The 2,000-franc hand that pops instead of breaking: ETH's ORCA and the repairable robot
ETH Zurich's open-source ORCA hand delivers 17-DoF dexterity for under 2,000 CHF, with poppable joints you reprint instead of replace. Why repairable wins.
StructureGS: when a Gaussian splat learns where the hinge is
StructureGS adds oriented bounding boxes to 3D Gaussian Splatting so scanned parts stay separable — why AEC teams should demand articulated deliverables.
Rehearsal without concrete: MIT's SceneSmith and the sim gap my fleet still pays for
MIT CSAIL's SceneSmith generates robot training worlds from text with GPT-5.2 agents. From the site floor: why your IFC model is the map robots actually need.
Teaching site robots to feel: touch inferred from cheap egocentric video
TouchAnything and the EgoTouch dataset infer contact and force from RGB video alone — what it means for teaching contact-rich site robots in AEC.
Asimo Retired, the G1 Shipped: What Japan's 80,000-Hour Robot Data Commons Bets On
Asimo retired unsold; the $16k Unitree G1 shipped. Inside Japan's 80,000-hour VLA data commons — and the openBIM-style bet on prototypes versus platforms.
A Site Robot's GPS Problem: Coarse-to-Fine Floorplan Localization Hits the Slab
A new arXiv method drops ray-matching for image-conditioned pose diffusion, holding sub-meter indoor localization on S3D and ZInD without lookup tables.
SONG rehearses the crowd before the robot walks your lobby
SONG's 3D Gaussian splatting simulator trains social-navigation robots on photoreal crowds — and turns every public splat-twin into a Swiss data-residency question.
Tesla built Optimus's factory before it showed the robot — and that inversion is the story
Tesla is converting a Fremont car line to build Optimus before revealing it — why the humanoid bottleneck moved from the demo to actuator supply and line rate.
One sparse factor, three passes: DiffPhD rewires the soft-body solver stack
DiffPhD, a GPU differentiable solver, stays stable at 100x stiffness contrast and runs 10x faster — bringing gradient-based form-finding to the AEC desk.
