Physical AI systems need real-world experience to work reliably, and a new wave of startups is building the data infrastructure to make that possible. Ropedia, a Singapore-based company, just raised $30 million across two funding rounds to expand its platform for collecting and preparing interaction data for robotics and embodied AI systems. The funding signals a shift in how the industry thinks about training robots: instead of relying on simulations or limited teleoperation, companies are investing heavily in capturing authentic human movement and task performance at scale.

Why Can't Robots Just Learn From Videos Like Humans Do?

The fundamental challenge facing embodied AI developers is that robots cannot learn physical tasks the way humans learn from observation alone. Ropedia's CEO and co-founder Zhaoxi Chen explained the gap bluntly: "A robot can't play baseball by watching a video any more than you could learn to ride a bike by reading about it. The robot must understand what it's like to grip a bat and know the timing it takes to hit a ball. That's the information Ropedia's technology provides". This insight captures why data collection has become a bottleneck. Traditional AI training relied on text and images scraped from the internet, but physical AI requires something fundamentally different: multimodal recordings that capture how humans actually interact with objects and environments.

Zhaoxi Chen

Ropedia's approach centers on a wearable device called HOMIE, which records first-person video, audio, depth information, hand tracking, eye gaze, body motion, and camera position simultaneously. Each data stream is timestamped so developers can link what a person perceives with what they do in real time. The company has already assembled a dataset called Xperience-10M containing 10 million interaction episodes and more than 10,000 hours of multimodal recordings, spanning billions of synchronized frames from video, depth sensors, motion capture, and inertial measurement units.

How Is Ropedia Solving the Data Collection Problem?

The startup's business model differs significantly from traditional data-labeling services. Rather than annotating third-party data, Ropedia generates its own source material, then processes, synchronizes, and structures it before licensing datasets to robotics companies. This closed-loop pipeline approach allows the company to maintain quality control and ensure that data is aligned with the needs of physical AI models. More importantly, Ropedia claims its method can reduce data-collection costs by as much as 50 times compared with traditional approaches, making large-scale training data more accessible to a broader range of developers.

The company's wearable-first strategy offers a practical advantage over robot-based data collection. HOMIE devices can be deployed across different environments and users in parallel, without requiring expensive robot hardware or being tied to specific robot formats. This scalability matters because embodied AI systems need diverse training data spanning multiple tasks, environments, and user types to generalize effectively. Ropedia has already served more than a dozen North American companies working in embodied AI and spatial intelligence, suggesting that demand is emerging among developers who need larger volumes of structured physical-world training data as robotics models move beyond controlled lab settings.

What Role Does Edge Computing Play in Real-Time Physical AI?

While data collection addresses the training problem, another critical challenge is enabling robots to process information and act in real time. PSYONIC's Ability Hand demonstrates how distributed computing architecture can solve this challenge. The prosthetic hand, which is used by both human prosthetic users and robotics organizations, combines tactile sensing, force feedback, and distributed Arm-based microcontrollers to deliver responsive control in under 200 milliseconds, faster than the average human blink.

The key innovation is that PSYONIC did not rely on a single centralized processor. Instead, multiple Arm-powered microcontrollers are distributed throughout the hand, each handling sensing, motor control, tactile feedback, and wireless communication. This architecture enables individual components to react in milliseconds while maintaining modularity for future development. More than 300 users rely on PSYONIC's technology as a prosthetic device, while more than 90 robotics organizations use the same platform to advance robotic manipulation and embodied AI development. The fact that the same hardware can serve both human and robotic users creates a feedback loop: real-world interaction data from prosthetic use can inform robotic learning, accelerating progress in both domains.

What Security Risks Emerge as Robots Enter Real-World Operations?

As embodied AI systems scale from research labs to factory floors and warehouses, security becomes a critical concern that extends far beyond protecting individual robots. ESET, a cybersecurity firm, has blocked more than 3,000 malicious AI skills since March 2026 and flagged a further 25,000 as suspicious. The company scanned nearly 800,000 unique AI skills, which are instructions that tell an AI agent how to perform tasks and interact with external systems, revealing how quickly the software infrastructure around AI agents is becoming a target.

Juraj Janosik, vice-president of artificial intelligence at ESET, emphasized that the risk profile changes fundamentally when software can influence physical action. "Embodied AI changes the risk profile because software can now influence physical action, and in some environments, human safety," he stated. The machine itself is rarely the point of entry for attackers. Instead, the softer targets surrounding it pose greater risk: stolen credentials, fleet management consoles, cloud services that machines report to, and update channels that push changes to every device on a site. An attacker who compromises these systems may be able to influence many devices at once, offering far greater return than targeting a single robot.

Steps to Secure Embodied AI Systems Across the Full Technology Stack

Protecting physical AI deployments requires a comprehensive approach that addresses multiple layers of the technology stack. Security experts recommend the following measures:

• Separate Safety from AI Decision-Making: Safety-critical functions such as emergency stops, collision avoidance, and operational limits should remain under deterministic control, with AI used only to optimize operations within those boundaries. The AI can decide how to complete a task, but it should never decide whether fundamental safety rules apply.
• Implement Multi-Sensor Verification: No single sensor should ever determine a physical action on its own. Camera data should be checked against Lidar (light detection and ranging), radar, telemetry information, environmental context, and expected behavior. When signals conflict, the system should recognize uncertainty and move into a safer operating state rather than act.
• Deploy Staged Rollouts and Continuous Monitoring: Every deployment should include staged model roll-outs, adversarial testing, rollback procedures, and continuous monitoring to keep behavior predictable as underlying systems change and models evolve.
• Protect the Full Supply Chain: Securing embodied AI means protecting the entire chain of interactions, including AI models, cloud services, software updates, application programming interfaces (APIs), sensors, third-party software, and fleet management platforms. Each dependency extends the chain of trust, and attackers consistently look for the weakest point rather than the most visible device.

Janosik noted that unpatched vulnerabilities, weak identity controls, and poorly governed software dependencies remain among the easiest routes into embodied AI deployments. This creates an uncomfortable trade-off for organizations that have built operational tools to manage fleets of machines. Centralized device visibility, staged model roll-outs, and over-the-air updates make continuous improvement possible across thousands of devices in the field, but from an attacker's perspective, they also provide leverage over every device at once.

At the Hyundai Motor Group Innovation Centre Singapore, the company's robotics smart factory embeds cyber security across its manufacturing environment, with a dedicated team continuously monitoring evolving risks while safety measures are designed into automation and robotics systems to protect people and equipment.

What Does This Mean for the Future of Physical AI?

The convergence of three trends is reshaping how embodied AI systems will be built and deployed. First, data infrastructure companies like Ropedia are making large-scale, high-quality training data more accessible and affordable. Second, edge computing architectures like those in PSYONIC's Ability Hand are enabling real-time responsiveness without relying on cloud connectivity. Third, security frameworks borrowed from aviation and industrial control systems are being adapted to ensure that autonomous systems remain safe and predictable as they become more capable.

Together, these developments suggest that physical AI is transitioning from a research curiosity to an operational technology. The companies investing in data infrastructure, distributed computing, and security are betting that the next wave of robotics will not be limited by hardware or algorithms, but by the quality and scale of real-world training data available to developers. As more robots enter factories, warehouses, and eventually homes, the infrastructure supporting them will become as important as the robots themselves.