A new architecture lets humanoid robots swap out their brains without forgetting how to carry a couch with you.
Researchers built a system called C2C that standardizes the physical meaning of signals passed between a robot's planning, coordination, and control layers, rather than standardizing the modules themselves. Task goals are represented as a geometrically verified path for the object being carried, a partner's movements are represented as physical state data that works no matter where it comes from, and learned coordination commands are bounded to an 11-dimension task space. Across nine transport scenarios, adaptive multi-agent reinforcement learning variants hit 77.1-82.1% mean success, compared with 56.5% for a scripted-partner reference. On physical Unitree G1 humanoids, the system reached 100% success moving objects through tight spaces and 80% success handling unusually long objects, and a three-carrier team of two G1s and one human validated the same setup on real hardware.
The headline numbers matter less than the plug-compatibility: swap the planner, the partner model, or the controller, and the rest of the stack keeps working because the physical contracts between layers don't change. That's the unglamorous problem that has kept human-robot teamwork brittle - most collaborative stacks are so tightly coupled that changing one part means re-tuning everything else.
Sim-to-real robotics papers have promised general-purpose robot teammates before. What's different here is a hardware demo with two humanoids and one human hauling an oversized object together, not just a simulation chart.