Physical AI's Promise Meets Reality: A Practical Strategy for Robot Companies

The hottest topic in the robotics industry today is undoubtedly "Physical AI." Just as ChatGPT broke down language barriers in the digital world, Robot Foundation Models (RFM)—artificial intelligence combined with physical entities that can judge and act autonomously—are now poised to reshape the industrial landscape.

However, a significant technological and economic gap remains between rosy expectations and the harsh realities of industrial settings. While the data-driven "brain" has advanced dramatically, connecting it to actual muscles and nerves presents challenges of an entirely different dimension. Customers seeking automation always raise the same concern: "Can we deploy Physical AI in our factory right now?" We must answer this question.

The Implications of Atlas's Roadmap

Hyundai Motor Group recently unveiled an ambitious industrial application roadmap utilizing Boston Dynamics' humanoid robot "Atlas." The plan centers on deploying the robot for parts sorting and sequencing at the Georgia Metaplant (HMGMA) in the United States by 2028, followed by parts assembly on the trim line by 2030.

The key point here is the difference in technical difficulty between 2028 and 2030. Sorting and sequencing occur on standardized pallets, but the trim line is the least automated process requiring the most sophisticated "tacit knowledge" from human workers.

Achieving this requires three evolutionary stages. Stage 1 is accurately identifying numerous objects in complex environments (Vision). Stage 2 is precise positioning and movement within a 0.1mm error margin (Motion). Stage 3 is manipulating objects flexibly and dexterously like humans using robot hands (Manipulation).

This raises a fundamental question. Giant corporations like Hyundai and Tesla can invest trillions of won in capital to design long-term goals three to four years out. But for most small and medium-sized AI and robotics companies lacking the massive capital and workforce to quench their "data thirst," an "all-in" strategy centered on VLA could be extremely risky.

Becoming consumed solely by training large models without immediate results, or focusing exclusively on low-probability robot hand manufacturing, risks technological isolation. What we need now is a "smart detour strategy" that captures practical benefits rather than blindly chasing large corporations. Small and medium robot companies especially need optimal path-setting for "tangible" short-to-medium-term results.

AI-Enabled Precision Control: The Emergence of 'System 0'

The "Helix 02" architecture recently announced by Silicon Valley's Figure AI demonstrates a significant paradigm shift. Until now, robot balance and stability relied on classical control algorithms written in over 100,000 lines of manual C++ code. But Figure AI boldly deleted this and added a neural network-based "System 0" layer with 10 million parameters.

This implements unconscious reactions that bypass cognitive judgment through AI—the core of physical resilience that ensures work continuity even under external shocks or on irregular surfaces. While System 2 handles logical reasoning and System 1 coordinates vision-motor functions, System 0 manages physical stability and contact dynamics (the "Last 0.1mm") in high-speed loops of 1,000 cycles per second.

When a robot walks while carrying a heavy tray, AI "intuitively" corrects minute changes in the center of gravity. The expansion of AI into precision control signals a paradigm shift: precision itself is the core of robot automation.

Achieving Results with Current Technology: The PSF Strategy

We must resist the temptation to view VLA as one giant "black box" that solves everything at once. Current multimodal LLMs and Vision Foundation Models (VFM) have matured enough to substantially handle the function of "seeing and understanding" industrial settings. The problem is that attempting to solve Stage 2 (precision motion) and Stage 3 (versatile manipulation) through data-centric learning alone delays validation due to insufficient training data.

High-difficulty manipulation using robot hands remains largely at the research stage, and achieving economic viability is not easy. So what practical detour strategy can deliver immediate results?

The answer is the "Physical Skill Foundation (PSF)" technique—clearly separating and combining Stage 1 (VFM) cognitive intelligence with Stage 2 (robotics-based precision control) physical intelligence.

Skill Combinations: A New Grammar for Industrial Automation

PSF involves securing numerous functional blocks of Actions that "see well" and "move well," then combining them to implement reusable "Skills." Skills are not simply listing actions in sequence. The essence of a skill is the "Behavior Tree"—the logical structure connecting actions.

For example, if a minor error occurs during a task of picking up a part and inserting it into a narrow gap, the robot doesn't simply stop. Following the behavior tree's failure recovery logic, it immediately returns to a "re-recognition" or "force-based fine adjustment" stage. Only when this "failure recognition and correction" function is integrated does an AI robot become a reliable piece of equipment that won't halt the factory line.

Using this technique, current AI technology can automate sufficiently complex industrial processes. Going forward, robots will understand their environment and tasks autonomously, and compose the skills and actions needed for work on their own. When that day comes, robots will be able to understand and execute tasks independently. This is our short-term goal.

Strip Away the Hype, Build Trust Through Proven Cases

The next two to three years will be the most dramatic inflection point in robotics industry history. What we need is "substantive Physical AI" that works immediately on-site and creates value.

Developing ambitious futuristic humanoids is important, but we must build modular skill foundations that can solve manufacturing pain points right now. Until the day comes when robots understand their environment and generate skills autonomously, we must achieve 100% reliability through flexible combinations of proven technologies. Ultimately, "reliability" equals "survival" and "competitiveness."

In the robotics field—where AI hallucinations can lead not just to wrong answers but to physical damage and human casualties—"the ability to recognize and recover from failures autonomously" will become a powerful weapon for leading global standards. Rather than being captivated by the glamour of intelligence, focusing on precision resilience proven in the field is the surest path for Korea's robotics industry to leap into the global leading group.