Robot Model SOTA Now Ships With a Nationality Qualifier
Two open weight robot models landed the same week, both citing benchmarks that no longer separate them. The differentiator has moved to adaptation and control.
#benchmarks
9 posts
humanoidsphysical-aiworld-modelssynthetic-dataros2planningsimulationgazebomanufacturingteslaagility-roboticsbenchmarks
Optimus starts, Cosmos opens the pipeline
Tesla begins building production Optimus units in Fremont, NVIDIA opens Cosmos blueprints and training tools, and PlanSys2 4.0.0 covers PDDL 1.2 through PDDL+.
manipulationphysical-airobot-securityros2open-source-hardwarebenchmarksdelivery-robotsprivacyreinforcement-learningtactile-sensing
Force models, backdoored planners, open arms
Eka bets on force instead of vision, a trigger word hijacks LLM robot planners, and open hardware gets a dual arm platform and five hands.
ros2open-sourcephysical-aiindustrial-robotssimulationhumanoidsmanipulationbenchmarksrobot-datavla-modelssim-to-real
Intrinsic Core, AWS stack, Atlas hands
Intrinsic open sources its industrial robot stack, AWS packages an open Physical AI pipeline, and a new benchmark shows frontier models failing at control.
humanoidsrobot-foundation-modelsbenchmarksimitation-learningindustrial-robotsphysical-airobot-datamanipulationautonomous-vehiclesrobot-safety
Frontier models fail humanoid control benchmark
Vals AI's Robo-bench scores frontier models under 2 out of 100 on humanoid control, Hitachi and FANUC partner on physical AI, and ABB ships new robots.
robot-safetyopen-source-hardwaremanipulationros2autonomous-vehiclessensor-fusionbenchmarksvla-modelslogistics
Mobile robot safety rules, open hand
A3 closes the user-side AMR safety standard, an open 15-DOF hand ships a ROS 2 driver, Tesla explains Robotaxi night limits, and benchmark means hide failures.
Helix 2.5, OpenArm, Arm capability levels
Figure's Helix 2.5 works in 30 homes it has never seen, an open humanoid arm ships CAD and ROS 2 packages, and Arm maps the physical AI stack.
Map-free drone nav, ARM cross-compiling
An open ROS 2 and PX4 stack flies a drone through urban terrain without a map, colcon gains ARM cross-compiling, and rosbag2 learns to change topics mid-recording.
Digit 5, OpenRobOps, Gazebo armhf cutoff
Agility's Digit 5 humanoid targets barrier-free factory work, InOrbit open-sources a fleet operations stack, and Gazebo ends 32-bit ARM binaries.