The New Automation Stack: Pioneering Trust, Safety, and Scale in Physical AI

At the IVS2026 conference held in Kyoto from July 1–3, 2026, a standout panel titled "Physical AI and the New Automation Stack" convened to address the next frontier of artificial intelligence. Moderated by Ryo Umezawa—a veteran serial entrepreneur, former country manager for Tinder and Halo, and global strategy leader at Vector—the discussion brought together three founders at the vanguard of physical automation: Junghee Ryu, Founder and CEO of RLWRLD; John Keh, Founder and CEO of Valtec Technologies; and Guido Cossu, Founder and Technology & Creative Director at Braid Technologies.

Together, the panelists explored how "Physical AI"—and its regional counterpart "Embodied AI"—is transitioning from controlled laboratory demonstrations into highly unpredictable real-world environments. The conversation focused on the practical mechanics of building trust, mitigating safety hazards, and navigating dense corporate bureaucracies to deploy automation at scale.

Meet the Founders: From Elite Air Force Operations to Robotics Pioneers

The panel brought together a diverse group of entrepreneurs whose technical foundations shape their unique approaches to physical automation:

• Junghee Ryu (RLWRLD): A veteran entrepreneur and investor in South Korea’s technology ecosystem, Ryu co-founded the prominent early-stage VC firm FuturePlay after selling his second venture to Intel in 2012. Following a personal battle with lymphoma four years ago, Ryu chose to return to active building, founding RLWRLD (pronounced "real world") to develop open-source robotics foundation models.

• John Keh (Valtec Technologies): Keh’s operational background includes serving as a drone operator in the United States Air Force, completing over 250 missions on platforms such as the Predator, Reaper, and Global Hawk. After transitioning to the private sector, he helped start the food-delivery platform Caviar (acquired by Square for $100 million) and ran business intelligence for Uber Eats across the United States and Canada before founding Valtec Technologies to bring aerial drone intelligence to commercial fishing.

• Guido Cossu (Braid Technologies): A theoretical physicist with a PhD in computational theoretical physics, Cossu spent many years conducting scientific research in Japan before co-founding Braid Technologies. Braid was established specifically to address advanced industrial-design bottlenecks, automating highly complex and time-consuming engineering workflows for physical hardware.

Defining the Physical Frontier

While large language models (LLMs) continue to dominate digital screens, the panel distinguished systems that generate digital output from systems that must perceive conditions, make decisions, and act safely in the physical world.

Ryu noted that while "Physical AI" is the marketing term popularized globally, the academic and East Asian research communities—particularly in Chinese markets—frequently use the term "Embodied AI." Ryu placed these systems in a clear hierarchy: Vision-Language-Action (VLA) models represent the direct descendants of Vision-Language Models (VLMs) and LLMs, designed specifically to translate perception into physical robotic actions.

Keh emphasized that Physical AI represents a paradigm shift from traditional, rigid industrial automation. Rather than operating as a "dumb robot" repeating a pre-programmed loop, a Physical AI system must actively interpret real-world conditions via onboard sensors and adjust its behavior dynamically.

The Architecture of Physical Trust: Managing Risk and Unpredictability

A central theme of the panel was the transition from idealized digital models to the messy realities of physical deployment. The founders outlined three distinct approaches intended to reduce unpredictable behavior and manage physical risk:

1. The Hybrid Control Framework (RLWRLD)

To bridge the gap between statistical AI models and deterministic safety requirements, RLWRLD implements a hybrid architecture. While Vision-Language-Action (VLA) models process high-level perception and spatial reasoning, they remain statistical engines prone to "hallucinations"—unpredictable outputs that could result in dangerous robotic movements.

To mitigate this, RLWRLD layers classical robotic control systems on top of the VLA output. This classical control layer acts as a safety filter, enforcing physical boundaries to prevent erratic or "weird, wacky" joint movements. To ground these actions in physical reality, the system integrates real-time measurements of physical forces, torques, and tactile pose data. During the session, Ryu shared a video demonstration illustrating these capabilities, showcasing a humanoid hand performing highly dexterous tasks, such as pouring liquids.

2. The Symbolic Physics Layer (Braid Technologies)

Braid Technologies approaches physical trust by bypassing statistical data models for its core design solver entirely. Cossu argued that equation- and constraint-based computation can avoid the type of unconstrained probabilistic output associated with generative models.

Rather than training neural networks on massive datasets to "re-learn" physics, Braid integrates established mathematical equations and physical laws directly into a deterministic symbolic layer. This provides mathematical guarantees that formally encoded structural and manufacturing constraints are satisfied without exception. Cossu noted that Braid works closely with heavy industrial clients to clearly define the boundaries of this automation, ensuring that every generated engineering part fits the client's precise fabrication requirements.

3. Human-in-the-Loop Validation and Rugged Engineering (Valtec Technologies)

For Valtec Technologies, the primary challenges are extreme environmental conditions and operational costs. Operating near the equator, Valtec’s maritime drones must survive severe ocean crosswinds during shipboard landings, salt spray, and intense UV degradation.

To manage physical risk, Valtec keeps ship operators directly in the loop. The system acts as a decision-support tool rather than an autonomous pilot: it gathers visual data from aerial drone cameras and integrates it via APIs with existing shipboard hardware, such as marine radar and sonar. The AI flags potential tuna catches, but the human captain makes the final decision on whether to reroute the vessel. This workflow prevents costly errors, as commercial vessels spend upwards of $10,000 per day on fuel and cannot afford to chase false detections.

Navigating Corporate Bureaucracy: Top-Down vs. Bottom-Up

For startups deploying physical AI, the engineering challenge is often matched by the difficulty of navigating enterprise sales. The founders shared contrasting strategies for working with large, established conglomerates:

• The High-Level Champion (Braid Technologies): Cossu warned that attempting to sell bottom-up through engineering teams is rarely successful. While engineers are eager to test new design software, they typically lack purchasing authority. To bypass months of bureaucratic delays, Braid focuses its sales efforts on executive leadership first, allowing top-down mandates to smooth the downstream integration process.

• The Parallel Approach (RLWRLD): Ryu outlined a "triple-threat" go-to-market strategy that runs bottom-up, top-down, and side-channel efforts simultaneously. RLWRLD engages directly with technical teams while utilizing corporate venture capital (CVC) arms to secure executive alignment. Additionally, they leverage global cloud and hardware partners—such as NVIDIA, AWS, and Microsoft—as trusted side-channel distributors to accelerate enterprise adoption.

• API and Visual Integration (Valtec Technologies): To overcome integration roadblocks with massive maritime hardware companies, Keh explained that Valtec remains highly flexible. While they pursue official API integrations with major radar manufacturers, they can also deploy immediately by using a simple, non-invasive workaround: pointing a drone-linked camera directly at a ship’s radar screen to extract and process data without altering the vessel's certified electronics.

The Strategic Moat and the Three-Year Golden Window

The panelists agreed that physical AI represents a highly capital-intensive sector, but one that offers incredibly sticky customer relationships compared to pure-play software. Once a physical AI platform is integrated into a customer's physical operations, the proprietary data feedback loops and vertical hardware integration create a powerful competitive moat.

Ryu concluded the session with a stark warning and a call to action for the East Asian ecosystem. He highlighted South Korea’s newly announced $4 trillion USD national AI infrastructure initiative—designed to secure water, electricity, memory fabrication, and AI factories—and referenced a statement by the Korean Minister of Science and Technology: the "golden window" for establishing foundational dominance in Physical AI is only three years.

With RLWRLD’s open-source RLDX-1 model currently outperforming benchmarks from prominent Western competitors, Ryu emphasized that East Asian startups have a rare, time-sensitive opportunity to lead the global physical automation stack.