Singapore's Ropedia Raises $30M to Build the Data Layer for Physical AI

While the world's robotics labs race to build ever-larger foundation models for machines that move, a young Singapore startup has quietly raised $30 million on a contrarian premise: the real bottleneck in physical AI is not the model. It is the data.

Ropedia, founded in late 2025 and headquartered in Singapore with an office in Mountain View, California, announced last week that it has closed $30 million in pre-Series A funding — an $8 million tranche disclosed in March followed by a newly revealed $22 million — from venture investors with AI, enterprise software and infrastructure experience across Southeast Asia, according to the company's announcement first reported by Tech Startups and Digital News Asia. Strategic partners in robotics and mobility also joined the round.

Wearables instead of teleoperation

Ropedia's founding team carries serious research pedigree. Chief executive Zhaoxi Chen comes from a background in 3D computer vision and multimodal AI, co-founder Fangzhou Hong worked on egocentric multimodal intelligence research at Meta, and chief scientist Ziwei Liu is an associate professor at Nanyang Technological University known for his work in computer vision and generative models.

Their core product is HOMIE, a head-mounted wearable that records first-person video, audio, depth, gaze direction, hand tracking, body movement and camera position, all with synchronized timestamps. Anyone can wear it, anywhere — which is precisely the point. Teleoperation rigs, the industry's default method for collecting robot training data, are constrained by expensive robot fleets and trained operators. Ropedia says its approach is roughly 50 times cheaper than a classic teleoperation setup, and it scales simply by handing out more devices.

"A robot can't play baseball by watching a video any more than you could learn to ride a bike by reading about it," Chen said in the announcement, arguing that internet-scraped footage will never teach machines the physics of real manipulation.

Ten million episodes and counting

The output of that collection effort is Xperience-10M, Ropedia's flagship dataset, which the company says already contains more than 10 million interaction episodes, over 10,000 hours of multimodal recordings and billions of synchronized video frames. Unlike vendors that relabel existing datasets, Ropedia generates raw data directly — a distinction that matters to robotics labs wary of provenance and licensing questions hanging over web-scraped corpora.

The pitch lands at an opportune moment. Embodied AI has become one of the hottest categories in venture capital this year, with humanoid makers and robot foundation model labs raising mega-rounds across the US and China. Every one of those models is starved for exactly the kind of high-fidelity human interaction data Ropedia is stockpiling — a picks-and-shovels position in a gold rush.

Singapore's growing physical AI bench

The round also underscores Singapore's emergence as a hub for the data and infrastructure side of robotics, complementing the city-state's strengths in research talent flowing out of NTU and NUS. Ropedia says the fresh capital will fund expanded data collection across Southeast Asia and North America, larger HOMIE fleet deployments, a stronger AI research platform and a growing engineering team in the United States.

The competitive question is whether proprietary human-experience data remains defensible as capture hardware commoditizes. For now, investors are betting that a two-continent head start and 10,000 hours of synchronized recordings are hard to replicate — and that whoever owns the data layer of physical AI will be paid by everyone building on top of it.