Track 1: Physical AI

Physical AI is AI that interacts with the physical world, such as robotics, autonomous vehicles, drones, industrial systems, and embedded intelligent devices. Unlike Cloud AI, Physical AI does not only generate outputs for a user or another software service. It must perceive real-world inputs from sensors and produce actions, control signals, or decisions that affect a machine operating in a physical environment.

Eligible projects may include one or more of the following:

• AI inference running on-device, on-robot, in-vehicle, or on nearby edge compute
• Applications that use real or simulated sensor data, such as camera, lidar, radar, IMU, GPS, audio, force, or other environmental signals
• Software that produces actions such as steering, braking, path planning, object manipulation, anomaly detection, navigation, alerting, or actuator control
• Systems combining AI with robotics or autonomy software stacks, such as ROS 2, simulation environments, middleware, or real-time control frameworks

This track includes AI workloads for:

• Robotics, such as robot perception, navigation, manipulation, motion control, voice interaction, and multi-robot coordination
• Automotive and autonomous systems, such as driver assistance, perception, planning, simulation, localization, safety monitoring, and in-vehicle AI pipelines
• Embedded and edge AI devices, such as smart cameras, sensor hubs, industrial controllers, drones, wearable devices, or intelligent appliances

Projects may run on Arm-based embedded, edge, or automotive platforms, including robotics SoCs, embedded Linux boards, microcontroller platforms, in-vehicle compute, or Arm-based edge servers used for local autonomy workloads.

Learning paths:

Robotics / middleware

• ROS 2 install guide
• Cyclone DDS install guide
• Robot Simulation and Reinforcement Learning Workflows on Arm
• Deploy multi-node Zenoh systems on Raspberry Pi  

Automotive / autonomous driving

• Deploy Open AD Kit containerized autonomous driving simulation on Arm Neoverse
• Prototype safety-critical isolation for autonomous driving systems on Arm Neoverse

Embedded / Edge AI

• Visualize Ethos-U NPU performance with ExecuTorch

Track 2: Cloud AI

Cloud AI is any kind of inference run on cloud compute infrastructure, such as AWS, Azure, or GCP, or on a dedicated server on-prem. Cloud AI is typically accessed as an API endpoint by other software, or with a user-accessible frontend. Optimizing AI for the cloud means looking at scalability and efficiency on the most cost-effective compute instances, which offer higher core count and larger memory than Physical or Mobile AI.

Eligible projects may include one or more of the following:

• AI Inference running on Arm-based compute, such as AWS Graviton, Microsoft Cobalt, GCP Axion, or Ampere based servers
• Applications optimized for CPU-based inference through processes of quantization or pruning that allow running more complex models on standard CPU-only cloud instances
• Using CPU-optimized frameworks such as ExecuTorch, LiteRT, and Llama.cpp to accelerate AI inference workloads on CPU-only instances

This track includes AI workloads for:

• Agentic workloads that combine multiple AI models, MCP servers, and integrations to complete complex tasks with little to no human intervention
• AI features like chat, summarization, classification, or transformation that can be scaled-out using the same cloud-native tools and techniques as the rest of the application stack

Learning paths:

• Deploy an AI Agent on Arm with llama.cpp and llama-cpp-agent using KleidiAI
• Deploy a Large Language Model (LLM) chatbot with llama.cpp using KleidiAI on Arm servers
• Deploy a LLM-based Vision Chatbot with PyTorch and Hugging Face Transformers on Google Axion processors
• Accelerate Whisper on Arm with Hugging Face Transformers
• Run vLLM inference with INT4 quantization on Arm servers
• Run a Large Language Model (LLM) chatbot with PyTorch using KleidiAI on Arm servers
• Run an LLM chatbot with rtp-llm on Arm-based servers
• Deploy SqueezeNet 1.0 INT8 model with ONNX Runtime on Azure Cobalt 100
• Deploy TensorFlow on Google Cloud C4A (Arm-based Axion VMs)
• Build a Multimodal Retail Restocking Assistant on Armv9 With MNN

Track 3: Mobile AI 

Mobile AI is AI that runs locally on Arm-powered client devices such as smartphones, tablets, and laptops. Unlike Cloud AI, Mobile AI performs inference on the device in the user’s hands, enabling low-latency, private, and offline-capable experiences for real applications.

Eligible projects may include one or more of the following:

• AI inference running fully on-device in a mobile or client application, such as text, vision, speech, or multimodal use cases
• Applications optimized for mobile constraints such as model size, memory use, responsiveness, battery awareness, offline use, or time to first token
• Projects using mobile or client inference frameworks such as ExecuTorch, ONNX Runtime, LiteRT / TensorFlow Lite, MediaPipe, or similar runtimes
• Cross-platform AI applications that run on Android, iOS, or Windows on Arm devices, as long as inference runs locally on Arm-powered hardware

 This track includes AI workloads for:

• Mobile phones and tablets, such as chat, summarization, translation, transcription, camera intelligence, and accessibility features
• Laptops and PCs, such as Windows on Arm systems running local copilots, creative tools, productivity features, or developer assistants
• Consumer on-device AI experiences that benefit from local execution, privacy, responsiveness, and reduced cloud dependence

Learning paths:

On-device LLM applications (Android)

• Build an Android chat app with Llama, KleidiAI, ExecuTorch, and XNNPACK 
• Build a customer support chatbot on Android with Llama and ExecuTorch 

Mobile AI frameworks and runtimes

• Run LLM inference on Android with KleidiAI, MediaPipe, and XNNPACK 

Performance and optimization on-device

• Measure LLM inference performance with KleidiAI and SME2 on Android