Researchers from the University of Southern California and VinMotion introduced Heave-Ho on November 20, 2026. They say the framework lets humanoid teams coordinate carrying and handover tasks through spoken leader-follower commands, without exchanging inter-robot network messages. The moving crew has a foreman, not a group chat. Source Source
The project reports scaling to 10 robots in simulation. Its real-world examples include four Unitree G1s carrying a table, plus smaller teams mixing robots and people. Source
One leader, several microphones
According to the researchers, a trained vision-language planner selects the next team skill. A leader speaks it, followers hear it through onboard microphones, and the robots begin on a shared schedule. Either a robot or a person can lead. Source
The researchers say they post-train a humanoid motion-tracking model using multi-agent reinforcement learning, with centralized training and decentralized execution. During execution, each robot uses its own local observations to handle physically coupled work, such as carrying a shared object. Source Source
The project page says the G1 control policy was trained with two identical robots and deployed without retraining in larger and mixed teams. EngineAIβs PM01 runs its own policy, trained using the same recipe. Source
The absence of inter-robot network messages doesnβt establish that the whole system runs offline. Source Source
What the big speed claim actually measures
With Heave-Hoβs planner choosing team skills, the researchers report 86% task success, versus 42% with a zero-shot baseline they label βGPT-6 Astra.β The full-system comparison used 10 trials on each of five tasks: real-world door and corridor tasks, plus simulated handover, table and sofa tasks. A human operator reached 98%. Source
The advertised roughly 120-fold speedup concerns planning decisions: the team reports 52.5 milliseconds per decision on a single RTX 4090, compared with an average of 6.46 seconds for the GPT baseline through its API. That is a local-GPU-versus-API latency comparison, not an established hardware-matched test. It does not put the furniture on fast-forward. Source
The researchers also announced HH-Bench, a simulation benchmark with 12 collaborative tasks for teams of two to 10 humanoids. Its tasks include table carrying, narrow-door passage, corridor navigation and box handovers, and can be combined into longer sequences. Source Source
The project page provides no payload measurements or workplace-safety evaluation to establish readiness for routine commercial work. Source
Comments