(C) Gemini Image

Imagine a humanoid robot stepping up to the back of a pickup truck, carefully navigating between boxes labeled with major logistics brands, and carrying a package straight to a building's doorstep. While a bit hesitant in its movements, the robot mimics human delivery tasks with remarkable precision—operating entirely autonomously without human remote control.

This scene comes from a recent demonstration video released by U.S. robotics startup RewardAI. Beyond delivery tasks, the robot seamlessly transitioned indoors, picking up items from stairs and mopping floors.

As the robotics industry accelerates toward human-like adaptability, tech giants and startups alike are tackling the ultimate hurdle: giving robots true "working intelligence" (ilmeori in Korean).

1. RewardAI: Capturing the Human Touch with Sensor Gloves

One of the greatest bottlenecks in robotics is a lack of training data. Unlike text-based generative AI that can easily scrape the internet, robots struggle to learn physical forces—such as the exact grip strength needed to hold an object—solely from videos.

To bypass the time and high costs of manual remote-control training, RewardAI developed the "OmniBody Hand," a sensor glove worn by humans. When a person wears the glove to open doors or pick up objects, it records finger movements, contact pressure, and surrounding environments. The company's proprietary AI model, OM-1, then translates this tactile data into autonomous actions.

However, industry experts note that the system's stability still needs to be rigorously tested in dynamic environments where object weights and positions constantly shift.

2. Google: Building "Thinking" Robots with Advanced Comprehension

While RewardAI focuses on manual dexterity, Google is tackling robot intelligence through advanced comprehension and reasoning.

Google’s humanoid, powered by its "Gemini Robotics 2" system, is designed to understand spoken instructions and execute multi-step tasks. In a recent demonstration, the robot successfully responded to a command to "put the watering can into the green bin under the shelf," visually identifying the objects and navigating the room accordingly.

Yet, linguistic comprehension does not automatically equate to flawless execution. Google acknowledges that its robots are still works in progress; internal benchmark tests showed success rates of 45.7% for picking up floor items and only 36% for screwing in lightbulbs.

3. NVIDIA: Scaling Up via Virtual Training Grounds

To solve the data scarcity problem from another angle, NVIDIA utilizes massive virtual simulations.

Using its Isaac Lab platform, NVIDIA trained thousands of small bipedal robots named "MicroDucks" to navigate slopes and obstacles simultaneously across 4,096 virtual environments. This approach allows robots to learn through trial and error without physical wear-and-tear or heavy hardware costs.

Nevertheless, virtual training has clear limitations. While physics engines can accurately simulate balance and walking, replicating tactile feedback—such as the exact texture of a surface or how an object deforms under pressure—remains deeply challenging in a virtual space. This explains why the industry continues to invest heavily in physical data collection tools like sensor gloves.

The Path to Robotics’ "ChatGPT Moment"

Ultimately, the true "ChatGPT moment" for humanoid robots will hinge on one critical metric: reliability in unfamiliar environments.

Wang Xinzhing, founder of Chinese humanoid developer Unitree, suggests that the turning point will arrive when a robot can successfully execute 80% of assigned tasks in 80% of previously unseen environments—a milestone he believes is just 1 to 2 years away.

However, industry observers emphasize a wide gap between curated demonstration videos and real-world execution. The ultimate test for the next generation of robotics will not be a single successful take, but the ability to handle hundreds of unpredictable real-world scenarios with minimal error.

[ⓒ Global Economic Times. 무단전재-재배포 금지]