MATE Lets Remote Operators Team Up to Train Humanoid Robots Virtually

Training humanoid robots to work together has long been bottlenecked by the sheer cost and logistics of physical multi-robot setups — multiple expensive machines, dedicated lab space, and constant resets between trials. A new paper posted to arXiv describes a system called MATE, or Multi-Agent virtual TEleoperation platform, designed to sidestep those constraints entirely by moving the collaboration into simulation.

According to the paper, MATE allows several operators in different locations to simultaneously pilot whole-body humanoid robots inside a single shared, physics-based virtual environment. Rather than eliminating the physical realism that makes collaboration data useful, the system preserves coupled physical interactions — how humanoids, objects, and the environment actually push, grip, and respond to one another — while removing the need for multiple real robots or operators working side by side in the same room.

Using this setup, the researchers built a dataset covering 24.1 hours of coordinated multi-humanoid behavior across 2,500 joint episodes. The dataset spans what the paper describes as five long-horizon tasks, including object handover, relay delivery, environment interaction, and cooperative transport — scenarios that require robots to time their actions relative to one another rather than act independently.

To help models learn efficiently from this interaction-heavy data, the team also introduced a technique called EAIS, short for Execution-Aligned Interaction Sampling. EAIS computes sampling signals within what the paper calls an execution-aligned prefix, and prioritizes moments in a demonstration that are either task-progressing or interaction-critical — effectively teaching learning algorithms to pay closer attention to the parts of a demo where robots are actually coordinating with each other, rather than treating every frame equally.

The researchers evaluated MATE-collected data using representative imitation learning methods as well as vision-language-action (VLA) policies, a class of models that map visual and language input directly to robot actions. Per the paper, experiments showed efficient data collection and effective policy learning across the collaboration tasks. Notably, the team also reported zero-shot transfer: policies trained purely on virtual demonstrations were able to operate on a physical humanoid robot without any additional real-world fine-tuning.

The work fits into a broader trend in humanoid robotics research, where teleoperation and simulation are increasingly used to generate the large volumes of embodied training data that whole-body manipulation and locomotion policies require. Most existing pipelines, the authors note, have focused on single-agent data collection; MATE specifically targets the harder problem of capturing genuine multi-robot coordination without the overhead of running several physical humanoids at once.

The paper, titled "MATE: Multi-Agent Virtual Teleoperation Platform for Humanoid Collaboration Data Collection," was submitted to arXiv's Robotics category (cs.RO) under identifier arXiv:2609.26520. A project page is referenced in the paper for additional details, though the specific URL was not resolvable from the abstract alone.