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.
Figures & tables
Figure 2: Overview of Heave-Ho . At the high level, the leader’s vision-language planner selects the next team skill from its camera and speaks it; every follower recognizes the command on board, and all robots start it on a shared skill schedule. At the low level, each robot maps the team skill to its own role, generates a reference conditioned on the object specification, and tracks it with a frozen foundation tracking model plus a residual policy shared by all robots. No network message is exchanged between robots.
Figure 3: Leader’s head-camera views over a task, each with the team skill chosen at that step: the narrow door in the real world (top) and the box handover in simulation (bottom).
Figure 4: The tasks of HH-Bench in simulation. Bottom: tasks can be chained into longer ones, here a stair carry, a placement on a dolly, and a handover of a second box to a third robot. Time sequences are in App. B.2 and videos on the project page .
Figure 5: Real-world experiments. The first four rows show two G1 robots over time, from left to right, labeled with the team skill being executed. The last row shows teams that differ from the training configuration, including a different humanoid and people. Faces are obscured for anonymity. Full time sequences of all teams are in App. C.2 and videos on the project page .
Figure 6: (a) Success rate of the decision makers, averaged over the five tasks of Tab. 1 , against decision latency. The shaded region is slower and less successful than ours. The operator selects the skills by hand and serves as the upper bound. (b) Success rate of the low-level controllers on each task of HH-Bench in simulation. Bars show the mean over three seeds, and error bars the standard deviation.
Real
Sim
Decision maker
Door
Corridor
Handover
Table
Sofa
Avg.
GPT-6 Astra (zero-shot)
3/10
4/10
7/10
4/10
3/10
42%
CHORUS [ 16 ]
0/10
0/10
6/10
0/10
0/10
12%
Heave-Ho (ours)
6/10
7/10
10/10
10/10
10/10
86%
Human Operator
9/10
10/10
10/10
10/10
10/10
98%
Table 1: Success rate of the full system with different decision makers, 10 trials per task. Door: narrow door; Corridor: L-shaped corridor; Handover: box handover from a truck; Table: carrying a table to the front of a sofa with four robots; Sofa: placing an L-shaped sofa on a target area.
Method
Success (%) ↑
Body (cm) ↓
Object (cm) ↓
Baselines
Pure SONIC [ 42 ]
0.22
58.24
83.20
Multi-HDMI [ 67 ]
28.56
68.57
66.92
Ablations
Heave-Ho w/o SONIC
16.78
92.69
89.53
Heave-Ho w/ independent actors
22.05
35.59
35.00
Table 2: Low-level controllers in simulation, averaged over the 12 tasks of HH-Bench , each with three seeds of 512 teams. Body and object errors are mean position errors over the full reference horizon, including frames after a failure. Per-task results are in App. E.1 .
Figure 7: Train with two, deploy with many. The policy trained with two G1 robots (top left) carries a table in simulation with teams of four, eight, and ten robots, without retraining.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Task
What it tests
Lift and carry
L-shaped corridor
steps along a turning path
Narrow door
rotating to fit through a door
Table carry
a large object shared by many robots
L-shaped sofa
an irregular, unevenly loaded object
Stair carry
partners at different heights
Appendix
Table 3: Tasks in HH-Bench , grouped by how the robots interact with the object.
Task
Team
Success
Narrow door
2 G1
9/10
L-shaped corridor
2 G1
10/10
Box handover
2 G1
10/10
Precise placement
2 G1
9/10
Narrow door
G1 + human
10/10
Stair carry
G1 + human
7/10
Appendix
Table 4: Real-world experiments. No network message is exchanged between robots.
Item
Value
Base model
Qwen3-VL-2B-Instruct
LoRA rank / alpha
32 / 64
LoRA targets (LLM)
q, k, v, o, gate, up, and down projections in every layer
LoRA targets (vision)
attention qkv and output projections and both MLP layers in the last 8 ViT blocks
Vision merger
fully trained
Trainable parameters
≈ 140 M (4.2 M in the ViT LoRA)
Appendix
Table 5: Training settings of the vision-language planner.
Item
Value
Physics / control time step
5 / 20 ms
Actor hidden widths
512, 256, 256
Critic hidden widths
512, 256, 128
Optimizer
Adam
Initial learning rate
3×10−4
Actor KL target
0.01
Appendix
Table 6: MAPPO settings for low-level policy post-training.
Success rate (%) ↑
Pure SONIC
Multi-HDMI
Heave-Ho w/o SONIC
Heave-Ho w/ indep. actors
Heave-Ho
L-shaped corridor
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
96.29 ± 0.68
Narrow door
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
0.39 ± 0.20
82.62 ± 1.55
Table carry
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
0.52 ± 0.30
97.01 ± 0.49
Sofa carry
0.00 ± 0.00
4.36 ± 0.88
0.00 ± 0.00
3.26 ± 0.69
93.36 ± 0.70
Stair carry
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
3.71 ± 1.09
100.00 ± 0.00
Precise placement
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
0.00 ± 0.00
97.66 ± 0.70
Appendix
Table 7: Per-task comparison of the low-level controllers in simulation. Each entry reports the mean over three seeds with 512 teams per seed. Superscripts show the sample standard deviation across seeds. Success rate is the percentage of teams completing a task. Body and object errors are mean position errors in centimeters over the full reference horizon, including frames after a failure.
Figure 8: Real-world time sequences of the teams shown as single frames in Fig. 5 , and of the team of three G1 robots and a person. Time runs from left to right. Faces are obscured for anonymity.
Figure 9: Time sequences of the lift-and-carry tasks in HH-Bench , in simulation. Time runs from left to right.
Figure 10: Time sequences of the handover, shoulder-carry, pull, and push tasks in HH-Bench , in simulation. Time runs from left to right.
Multi-robot collaboration allows robots to efficiently take on a wide range of tasks, from moving a couch through a doorway to assembling structures on a construction site. However, achieving such coordination in mobile multi-robot settings remains challenging: centralized methods conditioned on the combined observations of a team scale poorly with team size, and decentralized methods that train one policy per robot often require explicit alignment procedures or information sharing at inference time to overcome partial observability. Our key insight is that the visuomotor priors of pretrained vision-language-action (VLA) models should enable reactive, decentralized collaboration from each robot's local observations alone, without these inference-time assumptions. We propose CHORUS, a framework that adapts a single VLA backbone to control diverse, multi-robot teams. At inference time, each robot runs an independent copy of CHORUS, conditioned only on its own observations and a robot-identifying prompt. In real-world experiments including mobile tape measurement, library book handovers, and laundry basket lifting, CHORUS achieves a 64% point improvement over decentralized, from-scratch models, improves reactivity to teammate behavior by 40% points, and outperforms centralized baselines. Together, these results show that a shared VLA backbone is capable of achieving decentralized multi-robot collaboration, without per-robot policies or inter-robot communication at inference.
Multi-human cooperative object manipulation is a highly challenging topic in humanoid control. While existing methods can handle simple collaborative carrying tasks, they struggle with role-dependent, long-horizon, and sequential cooperative manipulation. To address this issue, we propose a hierarchical commander-executor framework that enables multiple agents to accomplish complex collaborative tasks in distinct ways. Specifically, a commander combines instructions and scene information for joint semantic and geometric planning. It then decomposes the task into role-specific subtasks and generates shared object-centric coordination plans, specifying execution timing, target positions, and manipulation actions for each agent. Conditioned on the commander's high-level strategy and each agent's local observations, the executor learns a low-level physics-based policy through a solo-to-cooperative curriculum to realize coordinated human-object interactions. The planned object trajectory provides a shared coordination target, while root positions and grasping instructions specify role-dependent execution cues. We evaluate our method on a variety of object manipulation tasks, demonstrating its ability to execute long-horizon tasks that combine different types of collaboration and to scale from two-agent to multi-agent cooperation. Extensive experiments further show that our approach achieves higher success rates compared to state-of-the-art methods. Code will be made publicly available at https://github.com/TJU-3DV/Co-Manipulation.
Many physical tasks in human environments require collaboration, from assisting a partner to jointly manipulating an object. Yet, existing humanoid benchmarks largely focus on single-humanoid skills and lack evaluation of multi-humanoid collaboration under egocentric visual observations. We introduce CoHuB (Collaborative Multi-Humanoid Benchmark), a simulation benchmark for multi-humanoid collaboration under egocentric visual observations. CoHuB provides 10 tasks, eight with two humanoids and two with three humanoids, spanning diverse collaboration patterns. We also provide synchronized demonstrations collected through a multi-operator VR teleoperation pipeline, in which each operator controls one humanoid from its egocentric view. Experiments with representative visuomotor policies reveal substantial challenges across different forms of coordinated perception and control. CoHuB provides a foundation for developing and evaluating multi-humanoid collaboration policies.
Hyunjin Park, Jebeom Chae, Minwoo Park +11
Yonsei University · Gwangju Institute of Science and Technology · University of California, Berkeley +1