USC Heave-Ho Scales Humanoid Teams to 10 Using Speech Signals
TL;DR
- USC's Heave-Ho hits 94% low-level control success in simulation versus ≤29% for baselines on coordinated humanoid loco-manipulation.
- Trained on two humanoids, the system scales to teams of up to ten and generalizes to heterogeneous G1+PM01 mixes without retraining.
- Its vision-language planner completes 43 of 50 long-horizon trials versus 21 for zero-shot GPT-6 Astra at roughly 120× lower per-decision latency.
A University of Southern California team behind Heave-Ho reports a decentralized framework for humanoid teamwork that hits 94% low-level control success in simulation versus ≤29% for baseline and ablated controllers, and completes 43 of 50 long-horizon trials against 21 for a zero-shot GPT-6 Astra planner at roughly 120× lower per-decision latency.
Trained on two identical humanoids, the system scales to teams of up to ten and, per the paper, "generalizes to heterogeneous teams without retraining," including mixes of G1 and PM01 units. In place of a network link, robots coordinate through a speech-based leader-follower mechanism driven by a vision-language planner, with each unit acting on local observations alone.
The authors frame the problem as needing "coordinated whole-body interactions, coherent long-horizon planning, and precise synchronization across robots," and introduce HH-Bench, a 12-task benchmark spanning team sizes of 2 to 10 that covers table carries, narrow doors, L-shaped corridors, box handovers and stair carries. The preprint lands the same day as NUS ForceHarness, another humanoid-manipulation result we tracked, continuing a run of papers positioning themselves against the GPT-6 Astra baseline.
Real-world experiments are described as showing "robust decentralized collaboration across challenging tasks." The abstract does not break out per-task real-world success rates.
Originally reported by huggingface.co
Read the original article →Original headline: Heave-Ho Decentralizes Multi-Humanoid Teamwork, Scales From Two Trained Robots to Teams of Ten With No Network Comms