Singapore-based Ropedia has launched HOMIE Gen2, a wearable multimodal capture system designed to turn real-world human activity into structured training data for robots, world models and embodied AI systems.


The product reflects a growing shift in physical AI: as models become more capable, access to diverse and accurately synchronized real-world experience is becoming a major bottleneck.


HOMIE Gen2 is designed to collect first-person data without requiring a robot, a motion-capture studio or fixed sensors installed in the surrounding environment. Ropedia argues that this approach can expand data collection from a small number of expensive robotic systems to a much larger network of human participants.


Capturing More Than Video

Ropedia distinguishes “experience” from conventional video. A recording becomes useful for physical-AI training only when visual information is aligned with movement, depth, pose, actions and changes in the environment.


The 380-gram HOMIE Gen2 headset includes four 1080p cameras, four digital MEMS microphones, an inertial measurement unit and support for up to 2 TB of local storage. Its four cameras each provide a wide field of view, enabling full 360-degree coverage around the wearer.


According to Ropedia's official specifications, the device supports up to approximately 90 minutes of recording and about four hours of standby time on its built-in battery. This differs from longer runtime figures reported in some secondary coverage; the official product specification is used here.


The system is designed to synchronize multiple data types into a shared spatial and temporal representation. Ropedia says its data engine can produce more than ten modalities, including stereo video, depth maps, hand and full-body key points, motion capture, task annotations, scene meshes and object tracking.


From Raw Capture to Training-Ready Data

The hardware is one part of a broader software stack. Ropedia's spatial foundation models process captured recordings into machine-readable annotations such as localization, depth estimation, hand-object interaction tracking and full-body motion capture.


The company offers the system in several configurations, ranging from the capture hardware alone to a full platform with fleet management, task distribution, recording review and data-pipeline tools.


This structure is intended to reduce the manual cleaning and alignment work normally required before physical-world data can enter model training.


HOMIE Gen2 follows Ropedia's Xperience-10M dataset, which contains multimodal recordings of human activity for robotics and embodied-AI research. The company describes the product as part of a “Human Experience Engine” that links data collection, automated processing and feedback from downstream models.


The Race to Scale Physical-AI Data

Robotics datasets are often constrained by a one-operator-to-one-robot collection model. Robots are expensive, difficult to move between environments and limited in number. A wearable system allows people to record tasks directly in homes, factories, hospitals and workplaces without waiting for a robotic platform to be available.


This does not eliminate the need for robot-generated data. Human demonstrations and robot trajectories capture different dynamics, and models still need to learn the physical constraints of specific machines. But wearable collection could provide a larger and more diverse source of first-person experience for world modeling, simulation and vision-language-action training.


Ropedia's bet is that the next phase of physical AI will depend not only on larger models, but on infrastructure capable of producing reliable real-world experience at scale.