How Physical AI and 6G power the next wave of robotics

What you should know:

• The evolution toward 6G and edge AI are shifting robotics from isolated, single‑robot autonomy to scalable, heterogeneous fleets, including humanoids, AMRs, drones and manipulators, working coherently toward shared industrial goals.
• Traditional 2D‑centric simulation limits real‑world autonomy. Qualcomm Technologies' innovation in XR technology can enable rich, 3D data capture from real environments, accelerating sim‑to‑real transfer, improving AI model robustness, and significantly enhancing reinforcement learning and fleet‑wide learning outcomes.
• By converging XR, advanced wireless and on‑device AI, robots can continuously learn from real deployments; share policies and world models across fleets; and deliver continuous, compounding value, positioning Qualcomm Technologies to lead the market for scalable, always‑learning robotic systems.

The next wave of robotics isn’t just about smarter individual machines — it's about AI-driven systems that can perceive, reason and act together in real time. As robotics evolves from standalone automation to coordinated, intelligent fleets, the network becomes a critical component. Reliable collaboration now demands ultra‑low latency, deterministic performance and built-in security, often in bandwidth-constrained, safety-critical environments where failure can be costly .

Robotics is an example of this, where we often see one of the most compelling and complex edge AI challenges. That complexity spans multiple dimensions:

• Embodied intelligence at scale: Full‑sized humanoids introduce massive degrees of freedom, especially in dexterous hands, paired with diverse sensor modalities to see, hear, feel and reason about real‑world physics.
• Long‑horizon reasoning: Robots must plan and execute complex manipulation tasks over extended timeframes in dynamic, unstructured environments.
• Data scarcity: Unlike digital AI, physical AI must generalize across skills with limited real‑world training data, making scalability and transfer learning essential.
• Operational deployment: AI must be delivered through a robust, deployment‑ready platform that supports fleet management, collaborative robotics and robot‑to‑infrastructure awareness, integrating sensing, compute and connectivity end‑to‑end.
• Power efficiency: All of this must operate within strict power envelopes to maximize uptime, enable efficient mechanical design and manage thermal constraints.

Together, these challenges signal a fundamental shift: robotics is no longer about optimizing isolated machines, but about building Physical AI systems that can operate, learn and scale reliably in real-world conditions. As robots move from independent operation to shared autonomy, intelligence can no longer remain confined to a single device. It must move — securely, consistently and in real time — across machines, infrastructure and environments.

This is where advanced wireless architectures become foundational. Emerging 6G and AI‑native connectivity models are designed not just to connect robots, but to coordinate intelligence — enabling deterministic, low‑latency collaboration at the edge, even under the most demanding conditions where safety, reliability and security are non-negotiable.

 

Always-learning robotic fleets on the factory floor: a concrete industrial scenario

One of the clearest expressions of connected Physical AI can be found on the factory floor, where robots must coordinate continuously to maintain throughput, safety and uptime. Consider an electronics assembly plant where throughput and uptime are critical:

• Humanoid robots powered by the Qualcomm Dragonwing IQ10 processor are designed to handle complex, fine-grained tasks: routing cables, inserting connectors, tightening fasteners and performing ergonomically heavy operations that are hard to automate with fixed tooling. This is supported by the following skills: high DOF manipulation, bimanual coordination, visual servoing (multicamera based visual features) and natural human–robot collaboration.
• Robotic arms mounted at workcells perform highspeed pick-and-place, singulation and precision placement from conveyor belts or totes, enabled by skills in object detection and segmentation, grasp synthesis, force aware insertion and recovery from mis-grasp or misalignment.
• Autonomous mobile robots (AMRs) orchestrate material flow, bringing parts to workcells, returning completed assemblies and dynamically rerouting around congestion, enabled by mapping and localization, multi‑robot route planning and collaborative obstacle avoidance skills.
• Inspection drones perform overhead inspection of lines, racks and safety zones, detecting anomalies, blocked lanes or equipment issues enabled by skills in 3D mapping, anomaly detection and semantic labeling of factory assets.​

Together, these agents interact as a coordinated fleet:

• Humanoids request resupply tasks directly to AMRs via low latency 5G/6G sidelink, exposing their upcoming part needs and current task backlog for just-in-time delivery.
• Arms and AMRs share object and location priors (SKU dimensions, bin locations, obstacle maps) via deterministic wireless, so when a bin is moved, everyone’s world model updates consistently.
• Drones feed high-level situational maps (blocked aisles, new temporary structures) that update fleet routing and manipulation strategies in near-real time.​

On top of this, safety critical motion controllers run on dedicated safety islands, while reactive behaviors execute on-device (local collision avoidance, short horizon planning), while long horizon scheduling and optimization can take place at the edge or cloud in a more hybrid architecture (line balancing, predictive maintenance, global routing).

To make this possible, robotic fleets must integrate:

• Sensor and collaborative planning for shared situational awareness.
• Cybersecurity and privacy first frameworks, including air-gapped operations.
• Continuous learning and hybrid AI to refine skills in real time.
• Power efficiency and low-latency operations, ensuring safety and precision.

This kind of always-learning fleet illustrates how Physical AI systems rely on connectivity not only for communication, but for shared intelligence — with emerging 6G architectures extending today’s capabilities toward more deterministic, AI-native coordination.

 

Dragonwing Robotics: the platform approach behind scalable Physical AI

Delivering this level of coordination requires more than advanced algorithms or powerful hardware in isolation. It demands system-level platforms designed to integrate sensing, perception, planning, control and learning — all under real-world constraints. And this is precisely what the Dragonwing Robotics Platform tackles.

Our vision is that every physical embodiment can become a continuously learning, intelligent robot and coordinated robotic fleets are one of the most concrete, near term realizations of that vision. Each robot, whether a humanoid, arm, AMR or drone, becomes a node in a larger, always-learning network, continuously improving through shared data, models and world understanding across the fleet.

At varying levels of complexity, the Dragonwing Robotics Platform focuses on core robotic “skills” that recur across industries:

• Singulation and pick-and-place
• Grasping, dexterous manipulation and tool use
• Recovery, error handling and repositioning
• Navigation, path planning and collaborative motion

These skills can be composed and reused across very different embodiments. For example, the Dragonwing IQ10 processor — architected as the “brain of the robot” for industrial AMRs and full-size humanoids — leverages the same policy backbone that can underlie a bin-picking arm, a humanoid doing kitting and an AMR performing pallet moves, with embodiment-specific adapters handled by the platform’s heterogeneous compute for sensing, perception, motion planning and control under tight power budgets.

Its high-performance predecessor, the Dragonwing IQ9 similarly powers sophisticated humanoids and mobile platforms with large language and vision models. Today, it's already operating in the field with OEM partners. Additional Dragonwing processors meet the compute requirements of service and consumer robots that perform comparatively simpler tasks. These demands — ultra-low latency, high reliability, offline support or intermittent connectivity, power efficiency — are brought on by edge-first AI architecture that align with Qualcomm Technologies’ deep expertise in heterogenous compute (CPU, GPU, NPU), energy-efficient AI inference, real-time perception and decision-making, and are the foundation of the five core pillars of the Dragonwing Robotics Platform.

Heterogeneous Compute

At the core is a high efficiency heterogeneous architecture codesigned for peak performance-per-watt. Innovative memory hierarchies and real-time safeguarding allow robots to make rapid, reliable decisions on the edge — keeping mission-critical systems responsive and efficient. The Dragonwing IQ10, for example, is powered by an 18-core Qualcomm Oryon CPU and dedicated NPU designed to deliver up to 700 TOPS to handle high-density AI workloads, with multimodal sensing (supports up to 20 cameras, lidar, radar) and advanced motion control capabilities for use in human-coexisting robotics.

Compound AI System

Robots must operate intelligently and safely in complex, mixed criticality environments. The Compound AI framework integrates three levels of cognition that we can think of in three operational layers:

• Reflexive sensing: Deterministic, low latency control at the microsecond level.
• Fast action and locomotion: Reactive intelligence and local decision making on device.
• Reasoning and long-term planning: Higher order inference and fleet coordination via edge and cloud.

Together, these layers embody the new era of connected autonomy: every robot can perceive, reason and act locally, yet contribute to the collective intelligence of the fleet. Additionally, these systems combine to enable perception, spatial reasoning, millimeter level dexterity and human aware collaboration, where perception meets cognition to unlock autonomous intelligence.

Physical AI MLOps

Continuous learning drives lasting capability. Qualcomm Technologies’ Physical AI MLOps framework is designed to support behavior cloning, reinforcement learning and world model-based methods to train, evaluate and deploy robotic behaviors safely. This creates an automated pipeline from data to deployment, accelerating iterative improvement across fleets. While teleoperation provides valuable demonstrations, it is inherently limited; data capture is resource intensive and constrained by operational scope.

Dragonwing aims to address this by leveraging world models that simulate physical dynamics and generate synthetic experiences, allowing robots to explore countless virtual variations of real-world tasks. When combined with reinforcement learning (RL), robots can refine their policies through simulated trial and error, discovering optimal behaviors under diverse conditions. This hybrid approach integrating real and synthetic data reduces training costs, enhances generalization and significantly accelerates continuous learning loops.

AI Flywheel

Building scalable Physical AI systems requires the same capabilities that enabled autonomy in other safety critical domains. Qualcomm Technologies has applied this approach in its automotive ADAS platforms — managing learning across millions of real world edge devices — and that foundation now directly informs how robotics can scale.

Building on this heritage, the AI Flywheel accelerates data collection, curation and annotation across distributed robotic fleets. Each interaction becomes feedback that strengthens a shared intelligence layer, creating a data network effect where every robot benefits from the experiences of the entire fleet, fueling faster model refinement and adaptability across industries.

Robotics is also where this flywheel becomes fully physical. By combining XR-based 3D simulation with Physical AI (“glass physical AI”), edge and cloud compute, and low-latency wireless connectivity, Qualcomm Technologies aims to enable continuous learning loops that bridge real-world operation and virtual training. The result is a self-reinforcing AI data flywheel that is designed to operate across four stages:

1. Deploy & Observe: Robots operate in the real world, capturing heterogeneous sensory data through on‑device compute.
2. Learn & Share: Insights, policies and models are shared across fleets via advanced connectivity, enriching a shared world model.
3. Simulate & Refine: XR-driven 3D environments train and test new skills at scale before redeployment.
4. Redeploy & Compound: Each iteration yields smarter fleets and greater operational value.

This cycle compounds value over time, allowing every participating robot, humanoid, manipulator or drone to continuously improve without centralized re‑engineering.

Dragonwing Robotics intends to embody this future: scalable, heterogeneous, always-learning fleets that transform isolated autonomy into connected intelligence. As 6G emerges and advanced XR matures, the boundary between simulation, learning and deployment is expected to blur. With our technology foundation, robotics will be positioned to move from tactical automation to strategic, self-optimizing intelligence across industries.

Deployment‑Ready Development Platform

Qualcomm Technologies aims to democratize robotics advancement through a comprehensive toolkit, including model zoos, domain specific foundation models, SDKs, reference designs and text-to-skill generation. Developers can quickly prototype, validate and deploy, dramatically lowering barriers while ensuring production readiness.

And by designing platforms that assume AI-native connectivity from the outset,  Dragonwing Robotics platforms systems are built with the intent to evolve alongside emerging 6G architectures, supporting distributed intelligence across devices, edge and cloud.

 

Why now: connectivity, AI and data finally align

These platform capabilities are not emerging in isolation — they coincide with a broader convergence across AI models, data, simulation and connectivity. Historically, locomotion has seen strong progress, but complex manipulation with high dexterity and many degrees of freedom remained fairly out of reach because classical AI architectures required enormous amounts of task-specific data, impractical for complex, multi-degree-of-freedom manipulation. Scaling high-dexterity robotic manipulation was thus limited by compute power and data diversity.

That’s changing. Foundation models, trained on large, multimodal datasets, enable a single model to generalize across tasks, adjusting its reasoning and control strategies dynamically. Combined with Qualcomm Technologies’ proven track record in delivering in Digital Chassis, Autonomous Driving and XR — technologies we have already deployed at scale in safety-critical domains — we now capture rich, 3D scene data from real environments, replacing traditional 2D-centric simulation. This approach is designed to support dramatic improvements in sim-to-real transfer, model robustness and reinforcement learning efficiency.

Crucially, connectivity underpins this transformation. Advanced 5G Sidelink and the forthcoming 6G architectures are designed to allow deterministic collaborative communication among robots, enabling realtime collaboration with low latency, even in bandwidth constrained or safety-critical settings with a cyber security layer. There are three key paradigms that 6G will aim to enable for Physical AI:

• Robot to infrastructure: For example, a robotic arm on the factory floor could offload compute to nearby infrastructure forming a shared local intelligence. It can access live data streams from overhead camera sensors to detect that items have fallen off a shelf and autonomously navigate to rearrange them. In another scenario, robots could upload newly captured data representing rare operational events to the cloud or edge server, enhancing model retraining and improving skill generalization for future similar tasks.
• Multi-robot collaboration: As an example, multi-robot collaboration in a retail and inventory setup enables autonomous robots to coordinate tasks like shelf scanning, stock replenishment and warehouse picking in parallel, improving speed and accuracy.
• In-Robot communication: Traditional robots rely heavily on intricate wiring to connect sensors, controllers and actuators. These cables can break under repeated motion or stress, leading to downtime and expensive maintenance cycles. Wireless in‑robot links enabled by 6G can reduce this physical complexity, improving reliability, flexibility and ease of maintenance.

Together, these shifts redefine what robotics can deliver — not just as automation, but as a strategic capability across industries.

Scaling Physical AI: From tactical automation to strategic intelligence

Building on its experience delivering real-time AI systems at the edge across multiple domains, Qualcomm Technologies now intends to extend its silicon-to-systems expertise into robotics. The goal is to enable safe, scalable and economically viable robotic systems that deliver real value: from warehouses and retail shelves to restaurants and homes.

Today, many robots still rely on scripted logic suited for structured environments. Dragonwing aims to reimagine this paradigm by combining advanced reasoning, adaptable autonomy and AI-native connectivity, enabling robots to act as autonomous agents in unstructured, dynamic settings while contributing to shared fleet intelligence. Achieving this requires a bottom-up design anchored in AI-based reasoning, safety and multidomain flexibility.

This transition — from isolated machines to coordinated Physical AI systems — defines the next phase of robotics and underpins our approach to AI at the edge. From intelligent silicon to adaptive AI systems, the Dragonwing Robotics Platform aims to transform tactical automation into strategic, self-optimizing intelligence — turning every robot into a learning agent, every fleet into a collective brain and every deployment into the next leap toward general purpose autonomy.

As AI shifts to real-time, agentic systems embedded at the edge, robotics are becoming a primary expression of intelligent action in the physical world, not scripted machines. Emerging 6G architectures can help to accelerate this shift, delivering ultra-low latency, deterministic connectivity and seamless cloud-edge coordination. Through shared data and learning, the Dragonwing Robotics Platform intends to connect individual robots into a collective network that evolves in real time. It aims to bridge digital intelligence with embodied action to create truly scalable physical AI.

At Mobile World Congress 2026, we will demonstrate how AI and connectivity come together to make this vision real — showcasing how Physical AI systems drive industrial transformation in the AI era.