← Paper Digest

Heave-Ho: Decentralized Long-Horizon Multi-Humanoid Collaboration

Ting-Hsuan Chen, Siheng Zhao, Han Jiang, Qi Pan, Louison Lu, Jiaheng Luo, Henghui Bao, Yiyu Chen, Quan Nguyen, Yue Wang

November 2026

Two G1s hand a box down from the bed of a truck.

TL;DR Heave-Ho lets teams of Unitree G1 humanoids carry shared objects with no network link between them. Each robot runs one shared residual RL policy on top of a frozen motion tracker, and a leader speaks the next team skill aloud. Real teams succeed in 92 of 100 trials, and in simulation the controller succeeds in 94% of episodes against at most 29% for baselines.

What they do

Tables, sofas and large boxes often need several humanoids lifting together. Centralized controllers grow with team size and need network links between robots. Heave-Ho is fully decentralized: each Unitree G1 runs the same policy on its own computer from its own observations, and the robots exchange no network messages. The authors also introduce HH-Bench, 12 collaborative tasks for teams of 2-10 humanoids.

How it works

  • Composable team skills. Lift, put down, steps and in-place rotations all start and end at two hub poses (standing empty-handed or holding the object), so any skill can follow any other. Reference motions come from retargeted videos of two people and from inverse kinematics, conditioned on object size and mass.
  • Residual multi-agent post-training. A frozen humanoid foundation tracker (SONIC) supplies balance and stepping. A residual actor shared by all robots learns the load sharing; it is trained with MAPPO and a centralized critic in Isaac simulation with two G1s. Partners affect each robot only through the object.
  • Planner and speech sync. A LoRA-tuned Qwen3-VL-2B picks the next skill from the leader’s head-camera image. The leader speaks it, followers recognize it by microphone, and all robots start on a coarse shared tick grid so acoustic delay does not shift the start. A person can follow or lead.

Results

  • Real world: 92 of 100 trials succeed across 10 team and task combinations; stair carry with a person is hardest (7/10). Teams unlike the training pair (4 G1s, G1s with people, an EngineAI PM01) succeed in 37 of 40 table carries.
  • Simulation: 94.41% average success over the 12 tasks, versus 28.56% for Multi-HDMI, 22.05% with independent actors and 0.22% for the frozen tracker alone.
  • Planner: 86% success on five long-horizon tasks, versus 42% for zero-shot GPT-6 (Astra), 12% for CHORUS and 98% for a human operator. One decision takes 52.5 ms, against 6.46 s for GPT-6 (Astra).

Limitations

Synchronization needs a reasonably quiet environment and reliable sound pickup. Training the planner for a new task needs teleoperated demonstrations. Teams of eight and ten robots were tested only in simulation, and real trials start from a nominal configuration, so real and simulated success rates are not directly comparable.

Videos

A robot leader speaks each team skill and both G1s carry the box in sync.
Two G1s carry a box to a target region marked on the floor.

Figures

From Figure 1: Decentralized humanoid teams completing long-horizon collaborative tasks. Each robot acts on its own o...
From Figure 1: Decentralized humanoid teams completing long-horizon collaborative tasks. Each robot acts on its own observations; the leader's planner chooses the next team skill and speaks it, and followers execute it in sync without any network communication.
From Figure 2: Overview. The leader's vision-language planner selects the next team skill from its camera and speaks ...
From Figure 2: Overview. The leader's vision-language planner selects the next team skill from its camera and speaks it; followers recognize it on board and all robots start it on a shared skill schedule. Each robot maps the skill to its own role, generates a reference conditioned on the object, and tracks it with a frozen foundation tracker plus a residual policy shared by all robots.
From Figure 6: (a) Success rate of the decision makers, averaged over five long-horizon tasks, against decision laten...
From Figure 6: (a) Success rate of the decision makers, averaged over five long-horizon tasks, against decision latency; a human operator selecting skills by hand is the upper bound. (b) Success rate of the low-level controllers on each HH-Bench task in simulation, mean and standard deviation over three seeds.
Abstract

Collaboration promises to extend humanoid capabilities beyond what individual robots can achieve. However, realizing this potential requires coordinated whole-body interactions, coherent long-horizon planning, and precise synchronization across robots. We introduce Heave-Ho, a decentralized framework for long-horizon collaborative multi-humanoid loco-manipulation. At the low level, we develop a multi-agent post-training framework that adapts a humanoid foundation tracking model to physically coupled interactions between humanoids, enabling coordinated decentralized control from local observations alone. At the high level, we train a vision-language planner and design a speech-based leader-follower coordination mechanism to synchronize decentralized execution without network communication. Trained with two identical humanoids, our system scales to teams of up to ten robots and generalizes to heterogeneous teams without retraining. To benchmark these capabilities, we introduce HH-Bench, comprising 12 tasks involving teams of 2 to 10 humanoids, including heterogeneous teams. Systematic evaluations show that Heave-Ho consistently outperforms baseline methods, while real-world experiments demonstrate robust decentralized collaboration across challenging tasks.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.