Physical AI / Robotics

The Physical AI / Robotics WG focuses on the transition physical AI from research labs to real-world industrial deployments. To bridge high-level cognitive models with microsecond motor control, systems must balance real-time reflexes with agentic reasoning through hardware-isolated workloads. Autonomy requires energy-efficient compute architectures operating within low power envelopes, complemented by hardware-anchored functional safety and zero-trust cybersecurity. Standardized open frameworks prevent vendor lock-in across heterogeneous silicon. In 2026, the working group will deliver an open reference architecture, a benchmark suite for reflex stability, and complete cloud-to-edge starter kits, accelerating safe and scalable robotic fleets globally.

Charter & Goals

Mission

The Physical AI / Robotics WG focuses on the transition physical AI from research labs to real-world industrial deployments. To bridge high-level cognitive models with microsecond motor control, systems must balance real-time reflexes with agentic reasoning through hardware-isolated workloads. Autonomy requires energy-efficient compute architectures operating within low power envelopes, complemented by hardware-anchored functional safety and zero-trust cybersecurity. Standardized open frameworks prevent vendor lock-in across heterogeneous silicon. In 2026, the working group will deliver an open reference architecture, a benchmark suite for reflex stability, and complete cloud-to-edge starter kits, accelerating safe and scalable robotic fleets globally.

Deliverables

Deliverable 0: The Open Physical AI Reference Architecture & Hardware Abstraction Profile (DRAFT)

Deliverable 1: Standardized Physical AI Benchmark Suite for Reflex & Safety Stability (“PAI Bake-OFF 2026”)

  • Lead Contributors: Johns Hopkins University, AWS, NXP, DeepX, ADI
  • Scope & Technical Objective:
    Move the industry past misleading “peak TOPS” marketing claims by establishing the first standardized, reproducible benchmark suite measuring edge silicon stability under concurrent cognitive and robotic reflex workloads.
  • Key Evaluation Vectors:
    • Perception Resilience Under Cognitive Load: Quantitative measurement of vision frame rates (FPS) and end-to-end perception latency while a local Vision-Language-Action (VLA) or Small Language Model (SLM) is simultaneously executing on the same system. Benchmarks mandate zero drop below the critical 15 FPS safety threshold.
    • Reflex Loop Determinism: Measuring worst-case execution time (WCET) and jitter on a closed-loop trajectory correction under synthetic memory and bus saturation stress tests (<5ms deterministic bound).
    • SWaP-C Efficiency Metrics: Standardized measurement of Joules-per-Physical-Task (J/PT), evaluating real-world work output per watt across mobile battery profiles.
  • Target Delivery: Open-source benchmark suite test harness and initial multi-silicon scoring published on GitHub and the Edge AI Foundation Wiki by mid-December 2026.

Deliverable 2: Cloud-to-Edge Robotics Blueprints & Starter Kits

  • Lead Contributors: Linux Foundation, AWS, NXP, Arduino, DEEPX
  • Scope & Technical Objective:
    Deliver a fully functional, open-source reference implementation demonstrating a complete cloud-to-edge lifecycle for physical robotics: from cloud-assisted simulation and model quantization down to sub-watt embedded edge deployment and real-time execution.
  • Key Implementation Components:
    • End-to-End Autonomous Demonstrator: A physical Autonomous Mobile Robot (AMR) / manipulator platform executing on-device multimodal perception (spatial vision + local speech commands via small models), dynamic obstacle avoidance, and compliant motor actuation completely offline without cloud dependency.
    • Cloud-to-Edge Fleet Continuum: Cloud pipeline leveraging AWS for model training, optimization, and secure OTA orchestration into containerized edge runtimes running on NXP and Arduino hardware.
    • Open Developer Starter Kit: A complete public repository featuring schematics, Bill of Materials (BOM), pre-trained quantized VLA/vision models, ROS 2 workspace nodes, and deployment manifests.
  • Target Delivery: Live physical demonstrator and hands-on workshop TBD – complete open-source code and hardware BOM released globally.

Meeting Cadence

The Working Group meets bi-weekly via Teams. Meeting minutes are published in the wiki within 48 hours. Ad-hoc deep dives are scheduled as needed and announced on the mailing list.

Decision Making

Decisions are made by rough consensus among active participants. For specification-level decisions, a minimum of two-thirds of the active members must approve. In case of deadlock, the WG Leads make the final call after documenting the dissenting views.

How to Contribute

  • Join the Working Group through the EDGE AI FOUNDATION member portal.
  • Attend a meeting to introduce yourself and learn about current workstreams.
  • Pick up an open task from the wiki or propose a new topic to the mailing list.
  • Submit pull requests to our GitHub repositories for code and documentation contributions.

Members

Davis Sawyer
Davis Sawyer Chair AI Product Marketing Manager NXP
Armaghan Ebrahimi
Armaghan Ebrahimi Business Development Manager NXP
Bill Podrasky
Bill Podrasky WW Head, IoT (Connected Products/Auto/Industrial/GenAI/Agentic AI/Physical AI) AWS
Evgeni Gousev
Evgeni Gousev VP of Technology Qualcomm
Jeremy Wallace
Jeremy Wallace Sr. Manager, IoT Solutions Architecture AWS
Marcello Majoncghi
Marcello Majoncghi Arduino CPO @ Qualcomm Arduino
Massimiliano Versace
Massimiliano Versace Global Head of Emergent AI Analog Devices
Tim Park
Tim Park Director Strategic Marketing DEEPX

Meeting Minutes

No meeting minutes published yet.

Output

No output published yet.

Articles