CoHuB: A Simulation Benchmark for Multi-Humanoid Collaboration

Hyunjin Park; Jebeom Chae; Minwoo Park; Sunghyun Park; Hanjun Yoo; Seoyeon Choi; Soochul Yoo; Joohwan Seo

cs.ROSep 28, 2026

CoHuB: A Simulation Benchmark for Multi-Humanoid Collaboration

Authors: Hyunjin Park, Jebeom Chae, Minwoo Park, Sunghyun Park, Hanjun Yoo, Seoyeon Choi, Soochul Yoo, Joohwan Seo, +6 more

Organizations: Yonsei University · Gwangju Institute of Science and Technology · University of California, Berkeley · Seoul National University

Abstract

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.

Figures & tables

Figure 1: Overview of CoHuB . CoHuB is a simulation benchmark for multi-humanoid collaboration, featuring 10 collaborative tasks with two or three humanoids, a multi-operator VR teleoperation pipeline for synchronized data collection, and phase-wise evaluation of policy performance.

Benchmark	Embodiment	# Agents	Whole- body	Collaboration	Human Teleop Demos
RLBench ( James et al., 2020 )	Manipulator	1	✗	✗	✗
LIBERO ( Liu et al., 2023 )	Manipulator	1	✗	✗	✓
RoboTwin 2.0 ( Chen et al., 2026b )	Manipulator	1	✗	✗	✗
BEHAVIOR-1K ( Li et al., 2022 )	Mobile manip.	1	✗	✗	✓
HumanoidBench ( Sferrazza et al., 2024 )	Humanoid	1	✓	✗	✗
SIMPLE ( Wei et al., 2026b )	Humanoid	1	✓	✗	✓

Table 1: Comparison of representative simulation-based robot policy learning benchmarks. Whole-body indicates whether tasks require coordinated locomotion and manipulation using the whole body. Collaboration indicates whether multiple embodied agents must collaborate to accomplish a shared task. Human teleop demos indicates whether demonstrations collected through human teleoperation are provided. ✓ yes, ✗ no.

Figure 2: The CoHuB task taxonomy. The 8 tasks with two humanoids are organized by base movement (horizontal axis) and physical coupling (vertical axis), with two tasks in each quadrant. Each task is illustrated by three key frames ordered from left to right.

Figure 3: CoHuB pipeline. The multi-operator teleoperation system records synchronized state trajectories, which are replayed offline to render RGB-D observations and construct the full demonstration dataset for policy training and evaluation.

	2 Humanoids									3 Humanoids
Method	Handover	Pouring	FrameHang	CoPouring	TrashCollection	CartService	CoCarry	TableAlign	Avg.	MoveHouse	BigTable	Avg.
Standard IL policies
ACT	65	1	44	9	1	3	61	39	27.9	9	80	44.5
DP	30	0	26	7	1	3	12	4	10.4	4	14	9.0
Multi-agent IL policies
LatentToM †	5	0	2	1	0	0	31	3	5.3	–	–	–

Table 2: Main benchmark results. Final task success rate (%) on the CoHuB task suite, evaluated from the beginning of each task. Blue and green rows denote per-task and multi-task training, respectively. Best result per task and best average are in bold . The shaded Avg. columns average across tasks and Task avg. across policies. Phase-wise results are provided in Appendix D .

Panel of Figure 4 — Figure 4: Phase-level evaluation results. Average success rate on non-collaborative and collaborative phases over the 8 tasks with two humanoids. Each phase is evaluated from its beginning state in each of the same 50 demonstrations. The red numbers indicate the relative drops.

		Standard IL		VLA
Obs. Scope	Policy	ACT	DP	GR00T N1.7
Local	Separate	27.9	10.4	14.5
Local	Shared	23.9	9.5	9.4
Global	Separate	25.6	19.1	12.3
Global	Shared	26.4	16.0	9.0

Table 3: Effect of observation scope and policy sharing. Average task success rate (%) over the 8 tasks with two humanoids across different observation scopes and policy-sharing strategies.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Simulator	Isaac Sim 6.0.1 with Isaac Lab
Physics / env. control	$200$ Hz / $50$ Hz
Robots	Two or three Unitree G1 humanoids, each with two Dex3-1 three-finger hands
Joint DoF	$43$ : legs $12$ , waist $3$ , arms $14$ , and hands $2\times 7$
Action space	$35$ : upper-body joint targets $31$ and base command $4$
Proprioceptive state	$43$ joint positions (rad)
Egocentric camera	RGB-D, $320\times 240$ ( $4{:}3$ ), $102^{\circ}$ horizontal field of view, $25^{\circ}$ downward tilt

Appendix

Table 4: Shared simulation and robot settings. Robot count depends on the task; action and state dimensions are per robot.

Panel of Figure 5 — Figure 5: Key frames of the CoHuB tasks with two humanoids and low base movement. Each row shows three key frames of one demonstration, ordered from left to right. (a)–(b) show tasks with low physical coupling, and (c)–(d) show tasks with high physical coupling. (a) Handover: one humanoid hands a bottle to its partner, which places it on the target. (b) Pouring: one humanoid pours a marble from a small cup into a glass held by its partner. (c) FrameHang: two humanoids jointly lift a large picture frame and hang it on a wall-mounted peg. (d) CoPouring: two humanoids lift a tub of apples together and tilt it to pour the apples into an empty crate.

Panel of Figure 6 — Figure 6: Key frames of the CoHuB tasks with two humanoids and high base movement. Each row shows three key frames of one demonstration, ordered from left to right. (a)–(b) show tasks with low physical coupling, and (c)–(d) show tasks with high physical coupling. (a) TrashCollection: one humanoid carries a bin to the end of a table, where its partner sweeps a bottle into it. (b) CartService: one humanoid pushes a trolley to a table, and its partner takes a wine bottle from the trolley and places it on the table. (c) CoCarry: two humanoids jointly carry a laundry basket to a shelf while keeping it balanced. (d) TableAlign: two humanoids carry a table across an office, rotate it during transport, and place it aligned with a chair.

Panel of Figure 7 — Figure 7: Key frames of the CoHuB tasks with three humanoids. Each row shows three key frames of one demonstration, ordered from left to right. (a) MoveHouse: one humanoid opens a door while the other two carry a table through the doorway and place it at the target location. (b) BigTable: all three humanoids jointly lift a large round table, carry it to the target location, and set it down together.

Task	Scene (source ID)	Task objects	Eval randomization
Handover	Kitchen Room/scene_084	Wine bottle; target pad	Bottle and target pad independently: $x,y\in[-0.025,0.025]$ .
Pouring	Kitchen Room/scene_085	Cup, glass, and marble	Cup and receiving glass independently: $x\in[-0.03,0.03]$ , $y\in[-0.06,0.06]$ , yaw $\pm 30^{\circ}$ ; marble position within the cup: $x,y\in[-0.006,0.006]$ .
FrameHang	Art gallery Room/scene_136	Framed painting; wall peg	Target peg and frame share $x\in[-0.05,0.05]$ and wall depth $y\in[-0.025,0.025]$ ; target peg $z\in[-0.04,0]$ ; extra frame offsets $x\in[-0.03,0.03]$ , $y\in[-0.01,0.01]$ .
CoPouring	Grocery store Room/scene_102	Tub of apples; receiving crate	Tub: $x\in[-0.02,0.02]$ , $y\in[-0.01,0.01]$ ; receiving crate (target), independently: $x\in[-0.06,0.06]$ , $y\in[-0.005,0.005]$ ; apples relative to tub: $x,y\in[-0.008,0.008]$ . Supporting stacks follow their containers.
TrashCollection	Science laboratory Room/scene_147	Bin; plastic bottle on bench	$+y$ toward Robot A. Bin: $x\in[-0.15,0.15]$ , $y\in[-0.15,-0.04]$ . Bottle: $x\in[-0.10,0.10]$ , $y\in[-0.04,0.12]$ .
CartService	Restaurant Room/scene_110	Service trolley; wine bottle	$+y$ toward Robot A. Trolley and its contents: $x\in[-0.10,0.10]$ , $y\in[-0.30,0.08]$ . Goal: $y\in[-0.16,-0.12]$ .

Appendix

Table 5: Task scenes, objects, and randomization. Source IDs refer to the source dataset before task-specific adaptation. Translation offsets are in meters.

Task	Scored phases and conditions
Handover	$p_{1}$ Grasp and lift: A’s right hand holds the bottle at least 3 cm above its counter resting height. $p_{2}$ Transfer: A’s and B’s right hands hold it together, followed by B’s right hand holding with both of A’s hands off, without an intervening all-hands release. $p_{3}$ Place: Released, stable, upright placement within $\pm 6$ cm of the target on each axis.
Pouring	$p_{1}$ Lift both vessels: Cup and glass are simultaneously at least 3 cm above the worktop. $p_{2}$ Tip: A holds the cup aloft and tilts it at least $60^{\circ}$ . $p_{3}$ Pour: The marble remains inside the glass continuously for 0.5 s, with the glass at least 3 cm above the worktop at completion. Returning the cup to the table is not required for success.
FrameHang	$p_{1}$ Grasp: Both robots hold the frame. $p_{2}$ Lift: They hold it together at least 10 cm above its spawn height. $p_{3}$ Hang: The peg passes through the frame’s loop while both robots hold the frame, which then remains threaded, stable, and released.
CoPouring	$p_{1}$ Joint grasp and lift: Both robots hold the tub at least 3 cm above its stack. $p_{2}$ Tip: They hold it together at a tilt of at least $25^{\circ}$ . $p_{3}$ Pour: All apples are inside the receiving crate, with both robots still holding the tub.
TrashCollection	$p_{1}$ Lift bin: A holds the bin at least 2 cm off the floor. $p_{2}$ Deliver bin: A holds it aloft within 25 cm of the bench-side carry point, with its rim no more than 1 cm above the bench top. $p_{3}$ Sweep: The bottle is inside the bin, with A still holding the bin.
CartService	$p_{1}$ Park: The trolley center reaches within 35 cm of the parking mark. $p_{2}$ Lift bottle: B lifts the bottle from the trolley positioned by A to at least 10 cm above its spawn height. $p_{3}$ Serve: B carries it within 15 cm of the goal in the horizontal plane, then releases it in a stable, upright placement within $\pm 6$ cm of the goal on each axis.

Appendix

Table 6: Success criteria of each phase. Phases follow

p_{1}\rightarrow p_{2}\rightarrow p_{3}

;

p_{3}

denotes task success. Distances refer to the manipulated object unless otherwise stated. For the tasks with two humanoids, the non-collaborative and collaborative phases used in Figure 4 are shown in blue and green, respectively.

Role	Instruction
Handover
Joint	Pass the bottle directly from Robot A’s right hand to Robot B’s right hand, then place it on the target.
Robot A	Pick up the bottle directly in front of you with your right hand and hand it to your partner’s right hand.
Robot B	Receive the bottle with your right hand, place it on the target directly in front of you, and release it.
Pouring
Joint	Pour the marble from the small cup into the glass the other robot is holding in the air.

Appendix

Table 7: Language instructions provided by the benchmark. Joint denotes the shared instruction. Robot A, B, and C denote the corresponding robot’s instruction.

Policy	Network initialization and training	Batch	Budget	Peak LR
Standard IL policies
ACT	Pretrained vision encoder; encoder finetuning and transformer training	$256$	$600$ epochs	$10^{-4}$
DP	From scratch; full policy training	$256$	$600$ epochs	$10^{-4}$
Multi-agent IL policies
LatentToM	From scratch; joint training of two policies	$32$	$1{,}000$ epochs	$10^{-4}$
GauDP	Pretrained Gaussian encoder; encoder finetuning, then diffusion policy training	$128$	$150$ epochs	$10^{-4}$

Appendix

Table 8: Baseline training configurations. Budgets refer to policy training; GauDP’s Gaussian encoder is trained separately as described below. Peak LR denotes the maximum learning rate. Band shading follows Table 2 : blue for the families trained separately per task, green for the VLA and WAM families trained jointly across tasks.

Method	Handover			Pouring			FrameHang			CoPouring
	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$
GPT-6 Astra	10	0	0	90	80	10	100	60	0	100	70	70

Appendix

Table 9: Astra phase-wise success rates (%) on tasks with two humanoids and low base movement. Phases are defined in Appendix B ;

p_{3}

is full-task success.

Method	TrashCollection			CartService			CoCarry			TableAlign
	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$
GPT-6 Astra	90	0	0	70	0	0	40	0	0	80	10	0

Appendix

Table 10: Astra phase-wise success rates (%) on tasks with two humanoids and high base movement. Phases are defined in Appendix B ;

p_{3}

is full-task success.

Method	Handover			Pouring			FrameHang			CoPouring
	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$
Standard IL policies
ACT	100	98	65	95	93	1	100	100	44	33	32	9
DP	87	64	30	38	38	0	76	73	26	21	18	7
Multi-agent IL policies
LatentToM	77	23	5	48	41	0	58	24	2	13	10	1

Appendix

Table 11: Phase-wise success rates (%) on tasks with two humanoids and low base movement. Phases are defined in Appendix B ;

p_{3}

is full-task success.

Method	TrashCollection			CartService			CoCarry			TableAlign
	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$
Standard IL policies
ACT	52	16	1	69	33	3	99	72	61	100	79	39
DP	32	4	1	66	9	3	96	31	12	93	34	4
Multi-agent IL policies
LatentToM	30	5	0	52	6	0	98	44	31	96	23	3

Appendix

Table 12: Phase-wise success rates (%) on tasks with two humanoids and high base movement. Phases are defined in Appendix B ;

p_{3}

is full-task success.

Method	MoveHouse			BigTable
	$p_{1}$	$p_{2}$	$p_{3}$	$p_{1}$	$p_{2}$	$p_{3}$
Standard IL policies
ACT	89	80	9	100	98	80
DP	73	57	4	98	50	14
Multi-agent IL policies
LatentToM †	–	–	–	–	–	–

Appendix

Table 13: Phase-wise success rates (%) on tasks with three humanoids. Phases are defined in Appendix B ;

p_{3}

is full-task success.

Panel of Figure 8 — Figure 8: Representative collaboration failures across the CoHuB task suite. Each pair shows two moments of one GR00T N1.7 rollout, earlier on the left. (a) Handover: the giver lets go before the receiver’s hand closes on the bottle, which drops between the two hands. (b) Pouring: the cup is tipped while the partner’s glass is not beneath it, so the marble misses the glass. (c) FrameHang: the robots lift the frame at mismatched heights and angles, so its loop misses the peg and the frame slips. (d) CoPouring: the robots fail to synchronize their grasps, leaving one handle ungrasped and causing the apples to spill outside the receiving crate. (e) TrashCollection: the bottle is swept off the laboratory table but misses the bin and falls to the floor. (f) CartService: the receiving robot carries the bottle to the goal table but walks past it instead of placing it. (g) CoCarry: one robot loses its grip on the basket, unbalancing the load and causing it to fall. (h) TableAlign: one robot loses its grip during transport, and the table falls outside the target region. (i) MoveHouse: the two robot carriers fail to coordinate their motion, causing the table to deviate from the doorway path and get stuck against the door frame. (j) BigTable: during transport, one robot loses its grip on the table, which tilts and never reaches the goal.

Explore similar work

Oct 1, 2026cs.RO

HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution

As robotic hardware and learning methods advance, humanoids need tools to perform tasks beyond their inherent physical limits. Successful tool use requires selecting a suitable tool and coordinating manipulation and, when needed, locomotion to complete the task. Existing benchmarks do not jointly evaluate these capabilities on a humanoid. We introduce HumanoidToolBench, an 18-task benchmark spanning three scenarios, three execution levels, and two tool-set modes, together with ToolBook, a dataset of 3.1k demonstrations collected in simulation and on a real Unitree G1. Evaluation of seven policies in simulation and three on the real robot reveals substantial gaps between selecting a suitable tool and completing the task. Focused GR00T N1.7 probes show reduced selection accuracy on unseen tools and continued task execution under unrelated instructions. Code and data are available at https://snu-pi.github.io/HumanoidToolBench/.

Kyochul Jang, Seohyeon Park, Ohchul Kwon +9

Seoul National University · University of Massachusetts Amherst · Google Research

Humanoid Controlling Tool Use Device

Ask PDF

Jun 30, 2026cs.MA

MECoBench: A Systematic Study of Multimodal Agent Collaboration in Embodied Environments

Recent multimodal large language models (MLLMs) have strong potential as embodied agents, but their ability to collaborate in visually grounded environments remains underexplored. To address this gap, we introduce MECoBench, a multimodal embodied cooperation benchmark with an evaluation platform spanning diverse real-world tasks, two cooperation structures, and three collaboration modes. Through extensive experiments across various MLLMs, we summarize three key findings: (i) Collaboration generally improves embodied task completion, but its benefits depend on balancing collaborative gains against coordination complexity. (ii) Communication is essential to collaboration gains, while the best collaboration mode depends on team size and model capability. (iii) Moreover, collaboration improves robustness under noisy priors and exploration conditions. Generally, MECoBench provides a systematic testbed for understanding the mechanisms and limits of multimodal embodied collaboration. Code and dataset are available at https://github.com/q-i-n-g/MECoBench.

Qingyun Liu, Jiwen Zhang, Jingyi Hu +2

1Fudan University · 3The Chinese University of Hong Kong · 2Shanghai Innovation Institute

Multimodal Agents Embodied Agents Multimodal Large Language Models Collaboration Cooperation Real-World Tasks+1

Ask PDF

Sep 22, 2026cs.RO

MATE: Multi-Agent Virtual Teleoperation Platform for Humanoid Collaboration Data Collection

Humanoid robots require diverse embodied experiences to acquire complex loco-manipulation and collaborative skills. However, existing humanoid data pipelines primarily focus on individual agents, while physical multi-robot collaboration remains difficult to scale due to costly hardware, dedicated spaces, and repeated resets. In this work, we introduce MATE, a Multi-Agent virtual TEleoperation platform for humanoid collaboration data collection that enables multiple geographically distributed operators to simultaneously control whole-body humanoids in a shared physics-based environment. MATE removes the need for multiple physical robots and co-located operation while preserving physically coupled interactions among humanoids, objects, and environments. Using MATE, we construct a multi-humanoid collaboration dataset comprising 24.1 hours of coordinated behavior across 2,500 joint episodes and five long-horizon tasks, including object handover, relay delivery, environment interaction, and cooperative transport. To improve learning from these interaction-rich demonstrations, we introduce EAIS, an Execution-Aligned Interaction Sampling strategy that computes sampling signals within an execution-aligned prefix and prioritizes task-progressing and interaction-critical behaviors. We evaluate MATE with representative imitation learning and vision-language-action policies across diverse collaboration tasks. Experiments demonstrate efficient data collection, effective policy learning, and zero-shot transfer from virtual demonstrations to a physical humanoid without real-world fine-tuning. Project page: https://yerik-yu.github.io/MATE/

Yichuan Yu, Youzhuo Wang, Yiming Ren +6

ShanghaiTech University · Nanyang Technological University

Teleoperation Multi-Robot Systems Strong Imitation Learning

Ask PDF