The annual Citigroup Research Robotics and Physical AI Leadership Summit concluded this Tuesday. The event brought together founders, investors, operators, and industry executives in the robotics field to assess the current state of "physical AI."

Citigroup analyst Heath Terry summarized the key takeaways in a note on Wednesday morning: the industry is moving from proof-of-concept to commercial deployment, but he also cautioned that scaling up robotics remains a significant challenge.

Terry told clients that "labor shortages, manufacturing reshoring, and a favorable regulatory environment are accelerating corporate demand," while "data scarcity, talent bottlenecks, battery life limitations, and high deployment costs remain key friction points."

Key Challenge: Data

Data scarcity was repeatedly highlighted as the central constraint during the summit.

Instawork noted at the event that even if the entire industry collects tens of millions of hours of real-world data by 2026, that volume would represent mere "basis points," not "percentage points," relative to the total data needed to achieve high-level robotic performance.

The analogy was clear: if the total required data is a swimming pool, the data collected so far is less than a bucket of water.

Unlike digital AI, where the foundational model of a large language model carries most of the value and can be rapidly replicated and deployed, the core value of physical AI lies in proprietary, task-specific data collected in real environments, combined with specialized hardware and safety certifications. This means that for nearly every new scenario or task, data accumulation must start almost from scratch.

Additionally, power supply, battery life, and chip architecture are becoming critical bottlenecks. Participants noted that existing semiconductor platforms are designed for data center workloads, not optimized for real-time, on-device inference on mobile platforms.

Winning Strategies: Start by Solving a Real Pain Point

The companies showing the fastest commercial progress at the summit—whether in humanoid robots, warehouse autonomous mobile robots (AMRs), autonomous trucks, or construction robots—all exhibited similar successful paths.

They start with a specific, high-pain-point labor problem rather than pursuing general-purpose capabilities.

They adopt a Robotics-as-a-Service (RaaS) model to lower the upfront procurement barrier for customers.

They prioritize safety and reliability over model complexity.

Terry believes that what has recently driven real returns on investment are specialized AMRs and systems from companies like Locus Robotics and Dexterity, rather than the much-hyped general-purpose humanoid robots.

While humanoid robots have attracted significant investment enthusiasm, near-term commercial returns are still primarily coming from these "specialized machines."

$20 Billion Inflows and Core Application Areas

Over the past two years, the physical AI sector has attracted approximately $20 billion in investment, with applications spanning warehousing, logistics, trucking, construction, aerospace, and defense.

Last week, BMW disclosed that upgraded humanoid robots are now walking and working on the production line at its Spartanburg plant in South Carolina.

On the demand side, several summit participants pointed out that logistics, warehousing, and automotive manufacturing are the core end-markets for current automation adoption. These scenarios share common characteristics: high-frequency, highly repetitive tasks suitable for robotic replacement.

Persistent labor market tightness and accelerated domestic manufacturing reshoring are two structural drivers pushing automation demand. Automation can increase production capacity, extend equipment uptime, and improve operational efficiency and precision, thereby supporting healthy return on investment.

The RaaS Model: A Key to the SME Market

High upfront costs have long been the biggest barrier to robot adoption for small and medium-sized enterprises (SMEs).

The emergence of the Robotics-as-a-Service (RaaS) model converts a one-time capital expenditure into a pay-per-use operating expense, significantly lowering the adoption barrier.

Terry specifically mentioned Symbotic's "Warehouse-as-a-Service" product (GreenBox/Exol), suggesting this model helps expand warehouse automation solutions to a broader customer base, including SMEs that were previously deterred by costs.

A Decade-Long Journey

Terry's final assessment is clear and direct: physical AI is a decade-long build-out and will not experience a rapid explosion like chatbots.

Advancements in AI and large language models, along with the growing richness of real-world and simulated data, are driving continuous technological iteration—hardware and software are becoming more deeply integrated, and systems are getting "smarter" with accumulated use. However, this process is gradual, not a leap.

Citigroup believes that long-term value will accrue to companies that master the data flywheel, solve real-world deployment problems, and meet the highest safety standards.