CoHuB Benchmark for Multi-Humanoid Collaboration Simulation

What happened

A research team has launched CoHuB, a collaborative multi-humanoid robot simulation benchmark. Built with an egocentric visual observation setup, the suite includes 10 tasks total — 8 for two-humanoid collaboration and 2 for three-humanoid collaboration — covering diverse collaborative modes. The benchmark comes paired with synchronized demonstration data collected via a multi-operator VR teleoperation workflow, where each operator controls a single humanoid robot from the robot's own first-person perspective. Tests on leading visuomotor policies show current models still face significant challenges in collaborative perception and control.

Key facts

Benchmark name

CoHuB (Collaborative Humanoid Benchmark)

Core observation setup

Egocentric visual observation

Task scale

10 total tasks: 8 dual-humanoid, 2 three-humanoid

Supporting resources

Synchronized demonstration data collected via multi-operator VR teleoperation

Background

Most existing humanoid robot evaluation benchmarks focus on single-robot manipulation skills, lacking assessments of multi-humanoid collaborative capability under egocentric vision. This leaves a mismatch with real-world requirements for multi-robot task completion in home and service scenarios.

Why it matters

For robotics researchers, CoHuB provides a standardized evaluation foundation for multi-humanoid collaborative algorithms, eliminating the cost of ad-hoc test environment setup. For developers, the accompanying teleoperation demonstration data can be directly used for collaborative policy training. For general users, as the technology matures, multiple service humanoids could eventually work together on complex daily tasks such as elderly care support and household chore coordination.

What to watch

Future developments to watch include real-world deployment progress of high-performance multi-humanoid collaborative control policies trained on this benchmark.