AgentAI's AI Robot Safety Verification Platform 'RoboGate' Featured on Global Academic Platforms

KOSDAQ-listed company AgentAI is gaining recognition for its technological prowess on the global academic stage with its independently developed AI robot safety verification platform, RoboGate.

On April 2, AgentAI announced that a research paper on RoboGate has been consecutively listed on major global academic platforms. According to the company, the paper has been registered on 'arXiv,' operated by Cornell University, and 'Zenodo,' run by the European Organization for Nuclear Research (CERN). It has also been submitted to 'SSRN,' a platform managed by the world-renowned academic publisher Elsevier, and is currently under review.

These platforms are recognized as representative academic repositories where major AI research, such as OpenAI's GPT and Google's AlphaFold, is first released, and where research data from international organizations like NASA and the World Health Organization (WHO) are accumulated.

Industry experts note that this case demonstrates RoboGate's technological standards conform to international benchmarks, as a domestic startup has successfully released a paper on major global academic platforms.

RoboGate is a platform that tests industrial AI robots in a virtual simulation environment before real-world deployment, allowing for preemptive checks of error possibilities under various extreme scenarios. It effectively serves as an "AI robot licensing exam."

Before actual factory deployment, RoboGate creates tens of thousands of scenarios in a "virtual factory" to identify potential issues such as misgrasping objects or collisions, thereby reducing risks in advance.

The AgentAI research team conducted a total of 50,000 simulations targeting robotic arms for industrial use and built a 'robot mistake dataset' by systematically accumulating failure cases across eight conditions, such as weight, friction, and speed.

Using this data, the team evaluated six major AI robot models (VLA), including NVIDIA's GR00T and Stanford University's OpenVLA. The results showed that all models recorded a 0% success rate in 68 extreme situation tests.

The company emphasized that even when the model size was expanded up to 260 times, there was no performance improvement, confirming that simply scaling up is not sufficient to resolve AI safety issues. This suggests that AI can be vulnerable to differences between training and real-world environments, underscoring the need for verification procedures like RoboGate before robot deployment.

AgentAI plans to accelerate the expansion of its business ecosystem centered around RoboGate. The company contributed benchmark code (PR #506) to NVIDIA's robot AI development platform, Isaac Lab, and is currently undergoing review to become an official NVIDIA software partner (ISV).

Furthermore, RoboGate is set to be integrated as an evaluator plugin in NVIDIA's "Physical AI Data Factory Blueprint," which is scheduled for release in April.

An AgentAI representative stated, "If NVIDIA is building the highway for AI robots, RoboGate will serve as the 'checkpoint' ensuring safety along that route. We aim to establish a safety verification system that must be passed before physical AI is deployed in real-world settings, setting a global standard."

Meanwhile, AgentAI plans to expand the number of AI models evaluated to more than 10, build a large-scale leaderboard, and pursue commercialization strategies such as a failure data subscription service (FaaS).

This content was produced with the assistance of AI translation services.