Heave-Ho Coordinates 10 Humanoid Teams Without Network Link

A University of Southern California research team has officially released Heave-Ho, a decentralized framework for multi-humanoid collaboration that coordinates teams of up to 10 robots using speech-based signaling rather than network communication. Officially launched on November 17, 2026, the system achieves 94% low-level control success in simulation compared to 29% or less for baseline controllers, and completes 43 of 50 long-horizon manipulation tasks versus 21 for a zero-shot GPT-6 Astra planner, operating at roughly 120 times lower per-decision latency.
Scaling From Two Trained Robots to Ten-Unit Teams
Heave-Ho trains on a pair of identical humanoid units and scales to teams of up to 10 without retraining. According to the research paper, the framework generalizes to heterogeneous robot configurations, including mixes of G1 and PM01 models. Each unit operates on local observations alone, with no network link between teammates. Instead, a leader-follower structure uses speech-based commands to synchronize actions across the team, enabling coordinated whole-body manipulation tasks that require precise timing and spatial awareness.
HH-Bench Benchmark and Task Coverage
The USC team introduced HH-Bench, a 12-task benchmark designed to evaluate decentralized humanoid teamwork. Tasks span team sizes from 2 to 10 robots and include scenarios such as table carries, narrow-door navigation, L-shaped corridor traversal, box handovers, and stair carries. The benchmark tests both low-level control precision and high-level planning coherence across long-horizon sequences. In simulation trials, Heave-Ho's vision-language planner completed 43 of 50 benchmark tasks, outperforming the GPT-6 Astra baseline by more than double while maintaining significantly lower computational overhead per decision cycle.
Whole-Body Control and Baseline Comparisons
The framework achieves 94% success on coordinated loco-manipulation tasks in simulation, a substantial improvement over baseline and ablated controllers that reached 29% or lower. The control architecture handles synchronized movements across multiple degrees of freedom, enabling robots to jointly lift, carry, and maneuver objects through constrained environments. The authors position Heave-Ho against both classical multi-agent planners and large language model-driven systems, emphasizing the efficiency gains from decentralized decision-making. The research lands on the same day as NUS's ForceHarness paper, continuing a recent trend of humanoid-manipulation studies benchmarking against GPT-6 Astra.
Real-World Deployment and Generalization
Real-world experiments demonstrate what the authors describe as robust decentralized collaboration across challenging tasks. The abstract does not report per-task success rates for physical trials, but the paper emphasizes generalization to heterogeneous teams and unstructured environments. The speech-based coordination mechanism allows robots to adapt leader-follower roles dynamically, reducing the need for pre-programmed choreography or centralized control infrastructure. This architecture positions Heave-Ho as a practical approach for environments where network reliability or bandwidth constraints limit traditional multi-agent communication protocols.
What This Means
Heave-Ho represents a shift toward decentralized multi-robot systems that prioritize local autonomy and speech-based coordination over network-dependent architectures. The framework's ability to scale from two trained units to 10 heterogeneous robots without retraining suggests potential for rapid deployment in logistics, construction, and disaster-response scenarios. By outperforming zero-shot large language model planners at a fraction of the latency, the system offers a blueprint for real-time humanoid teamwork that balances planning sophistication with computational efficiency. The 12-task benchmark provides a reproducible standard for future research in coordinated humanoid manipulation.
on Emergent today





