Heave-Ho: Decentralized Long-Horizon Multi-Humanoid Collaboration
Trained with two identical humanoids, Heave-Ho scales to teams of up to ten and mixed robot teams without retraining.
Each robot controls its movements using local observations, while a vision-language planner handles longer sequences of actions. A speech-based leader-follower mechanism synchronizes execution without network communication. The authors introduce HH-Bench, a benchmark of 12 collaborative tasks, and report consistently stronger performance than baseline methods. They also report robust collaboration in real-world experiments.
HF Daily Papers' note
Each robot controls its movements using local observations, while a vision-language planner handles longer sequences of actions. A speech-based leader-follower mechanism synchronizes execution without network communication. The authors introduce HH-Bench, a benchmark of 12 collaborative tasks, and report consistently stronger performance than baseline methods. They also report robust collaboration in real-world experiments.
HF Daily Papers' note
score 5