Executive Summary
Physical Artificial Intelligence (AI) is emerging as a major growth opportunity as Robotics Foundation Models (RFMs), Reinforcement Learning (RL), and multimodal AI expand what robots can perceive, learn, and automate. However, deployment robustness and a lack of high-quality, task-specific, real-world data present major barriers to scaling robotics deployments from Proof-of-Concept (PoC) to full-fledged operation. Unlike Large Language Models (LLMs), robotics cannot rely on Internet-scale datasets; it must build a continuous data flywheel from human demonstrations, real-world deployments, simulation, and RL. As battery-powered robots continuously learn and refine skills, scaling Physical AI will also depend on low-power inferencing and distributed intelligence across heterogeneous, hierarchical compute architectures, enabling decision-making at the edge rather than solely in a central robot controller. This creates growing demand for Edge AI, Extended Reality (XR)-enabled data capture, sensor fusion, and deployment infrastructure. Qualcomm is well positioned to support this ecosystem through its capabilities across automotive AI, XR, Internet of Things (IoT), edge computing, Dragonwing advanced Systems-on-Chip (SoCs), and connected intelligent systems, helping enable the next phase of scalable Physical AI.
The robotics industry is approaching an inflection point. While global automotive sales recently surpassed 100 million units annually, commercial robotics deployments remain comparatively modest at approximately 1.5 million systems worldwide. Yet enthusiasm surrounding robotics has never been greater. Advances in transformer architecture, RFMs, RL, diffusion policy, and multimodal AI have dramatically expanded what robots can perceive, understand, and accomplish. These technologies allow robots to transition from rigid, single-task, systems to generalist machines capable of operating in dynamic and unstructured environments. Venture capital investment continues to accelerate. Governments are encouraging the reshoring of manufacturing capacity, and a growing number of automotive manufacturers are exploring humanoid robots as a natural extension of their expertise in intelligent machines. Physical AI is rapidly emerging as one of the most important growth opportunities across the broader automation market.
RFMs, with the support of RL, have been proven to dramatically extend the range of tasks that can be automated by robots. Vision-Language-Action (VLA) models and world models (which innovators now combine to create World Action Models (WAMs)) have been successfully applied to previously un-automatable tasks, including manipulating deformable objects such as fabrics, clothing, plastic bags, wires, and even food; models can rapidly generalize across tasks; and swap between tasks and policies creating generalized autonomation. For these reasons, ABI Research forecasts a US$150 billion market for RFMs by 2036.
Much of the discussion surrounding Physical AI has focused on model architecture and the advanced SoC needed to run them. However, the industry's biggest challenge is not necessarily developing the next RFM. Rather, it is creating the infrastructure needed to continuously train, improve, deploy, and scale intelligent robotic systems. Robotics’ biggest challenge is data.
LLMs benefited from access to Internet-scale datasets accumulated over decades. RFMs have no equivalent resource. Robots must learn from interactions with the physical world, where collecting task-specific real-world data is expensive, time-consuming, and often difficult to scale. Open-source datasets and third-party repositories provide valuable starting points, but they frequently lack the domain-specific and task-specific information required for robots to perform real-world work reliably. As robotic systems move from controlled demonstrations to commercial deployments, the ability to collect, curate, and utilize physical-world data is emerging as a primary differentiator.
Robotics presents many of the same foundational challenges as autonomous driving, while introducing additional complexity through physical interaction with the real world. Vehicles operate within established road rules, while robots must contend directly with physics: manipulating objects, handling unpredictable environments, and managing dexterity, force control, and context.
Historically, simulation has been viewed as a solution to this challenge. Simulators allow developers to generate synthetic data, test robot policies, and accelerate training. Yet the industry is increasingly recognizing the limitations of simulation-only approaches. While visual perception can often be modeled effectively, accurately reproducing real-world physics remains difficult. Force feedback, contact dynamics, object manipulation, and dexterous motion continue to create significant gaps between simulated environments and real-world deployments. The Simulation-to-Real (Sim2Real) problem remains one of the most persistent barriers to scaling Physical AI and advancing RFMs.
Consequently, the industry is beginning to shift toward a more comprehensive approach centered around the AI data flywheel. Rather than relying exclusively on synthetic data, robotics developers are increasingly seeking methods to collect large quantities of real-world operational data, continuously refine models, and rapidly redeploy improvements back into the field. RL is emerging as a critical component of this process. Human demonstrations can provide an initial, ground-truth robot policy, but RL enables systems to refine those behaviors through repeated interaction with both simulated and real-world environments. Over time, robots can optimize task execution, adapt to edge cases, improve manipulation performance, and generalize beyond the scenarios present within the original training data. This creates a continuous learning cycle whereby data collected from the field improves models, improved models generate better behaviors, and new operational experience is fed back into subsequent training cycles.
One of the most promising developments within this emerging ecosystem is human demonstration learning, enabled through XR smart glasses and spatial computing platforms. While third-party datasets and open-source repositories provide a useful starting point, they often lack the task-specific fidelity required for real-world robotic deployments. Robotics systems must understand not only what action to perform, but how to perform it under varying physical conditions. Subtle differences in object orientation, force application, environmental conditions, and operator behavior can significantly impact task execution. As a result, the industry is increasingly looking beyond generic datasets toward scalable methods of capturing high-quality real-world data for training RFMs and other Physical AI strategies.
A compelling approach involves using lightweight XR smart glasses and spatial computing technologies to capture human demonstrations directly from skilled operators. Rather than relying exclusively on teleoperation, which requires a robot and trained operator in the loop, a worker can perform a task naturally while vision systems, hand tracking, and spatial sensors capture gaze, movement, object interactions, and task sequencing. These data can then be curated and transformed into ground-truth robot policies. RL then allows robots to transform these initial demonstrations into generalized behaviors capable of supporting a broad range of applications across humanoid robots, industrial robots, and Collaborative Robots (cobots). Qualcomm's experience across XR, Edge AI, hand tracking, and intelligent sensing technologies positions the company well to support this emerging AI data flywheel.
The importance of Edge AI extends beyond data collection. As robotic deployments scale, cloud-only approaches become increasingly impractical due to latency, connectivity, privacy, and cost considerations. Many robotic systems, including industrial robots, Autonomous Mobile Robots (AMRs), drones, and future humanoids, must operate in real time while maintaining low power consumption. This places increasing emphasis on highly efficient on-device AI inference and distributed computing architectures that push intelligence closer to sensors and endpoints.
These requirements create strong overlap between robotics and automotive technologies. Humanoid robots and next-generation vehicles increasingly share foundations in sensor fusion, AI acceleration, battery management, perception software, edge computing, foundation models, safety and security architectures and requirements, and continuous data collection.
This convergence gives automotive AI specialists a clear advantage. The industry’s experience in perception, decision-making, simulation, sensor processing, and scaled deployment is increasingly relevant to robotics developers seeking to commercialize Physical AI.
For Qualcomm, this convergence is particularly significant. The company's experience spans automotive platforms, IoT devices, XR technologies, Edge AI, and connected intelligent systems. Rather than focusing exclusively on a robot's "brain," Qualcomm is helping enable the wider Physical AI ecosystem: hardware platforms, software frameworks, AI toolchains, data pipelines, and deployment infrastructure that support continuous learning throughout a robot's lifecycle.
The near-term opportunities are especially compelling within manufacturing, logistics, and industrial environments. These settings provide structured, repeatable workflows where robotic systems can be trained, refined, and validated more effectively than in unpredictable consumer environments. At the same time, labor shortages, reshoring initiatives, and productivity demands continue to encourage greater automation across industrial sectors. Robots are increasingly being viewed not as replacements for workers, but as tools that can augment human capabilities and address difficult, dangerous, or undesirable tasks.
Ultimately, the future of robotics may not be defined by any single model architecture. New architectures will continue to emerge, and today's leading approaches may be superseded within months. The more durable competitive advantage will likely come from the infrastructure surrounding these models: the ability to collect data, train systems efficiently, deploy intelligence to the edge, and continuously improve performance through closed-loop learning.
The companies that succeed in building this data flywheel will help define the next era of Physical AI. By bringing together expertise in automotive AI, edge computing, XR, IoT, and intelligent connected systems, Qualcomm is positioning itself to play an important role in enabling that future. The next phase of robotics will not simply be about smarter robots—it will be about creating the ecosystem that allows robotic intelligence to continuously learn, scale, and evolve.
Related Research:
Physical AI Foundation Models Tracker
Dimitris Mavrakis
Larbi Belkhit