Kevin Weekly, Ph.D. · Physical AI
Physical AI — from research to shipped product. Perception, robot autonomy, on-device inference, multi-sensor fusion, and the safety-critical embedded software that lets a learned model actually act in the world.
I build systems where machine learning has to survive contact with the real world — running on the robot, on the wearable, on the drone, under hard real-time and safety constraints. Fifteen years across perception, SLAM, autonomy stacks, edge inference, sensor fusion, and the embedded software that holds it all together. Berkeley EECS Ph.D. in estimation; founder of an R&D firm; named inventor on granted and published patents in perception, robotics, and wearables.
Why this matters for an AI lab or a robotics product team
The hard part of physical AI isn't the model — it's everything around it: real-time constraints, messy sensors, hardware budgets, and a high bar for safety. That's the work I do.
Perception and autonomy that run on the robot — under hard real-time, on the compute you actually have, fusing sensors that lie to you, and closing the loop down to motors and brakes. I've built and led that work across autonomous vehicles, industrial mobile robots, and machines that shipped to customers.
Getting a network inside a power, latency, and memory budget is its own discipline: quantizing and hand-optimizing kernels for DSPs and NPUs, wiring inference into an RTOS, squeezing biosignal algorithms onto a wearable. The unglamorous half of ML — and the half that decides whether the model ever leaves the laptop.
Before a learned system is allowed to act in the physical world, you have to bound what it can do — and prove it. Confinement and functional safety, worst-case execution timing, verification frameworks, and a patent portfolio on safety-critical DSLs, hybrid safety verification, and AI agents for automated code validation.
I founded an R&D firm and grew it to $1.7M in revenue across 20-plus delivered projects, fielded a qualified entry in the XPRIZE Autonomous Wildfire Response competition, and have led every team I've been on. The through-line: taking systems all the way from research to something that works in the field.
Selected work
Ordered roughly by relevance to embodied / physical AI. Many were delivered under NDA — happy to go deeper in conversation.
Full vehicle autonomy stack: CNN-based perception and VLM workloads running on NVIDIA Jetson, multi-sensor fusion, real-time propulsion and steering control, a custom shield PCB, WebRTC video teleoperation, and a browser-based command-and-monitoring interface.
GridNet-style speech-enhancement network (STFT → CNN/LSTM → iSTFT) deployed to Qualcomm Hexagon, hand-optimized with HVX SIMD intrinsics and wired into the QuRTOS kernel for real-time multi-channel processing. The unglamorous half of ML: making it run within the power, latency, and memory budget.
Drone-mounted multi-camera fire-detection system and a drone-based gimbaled fire-suppression platform — fielded as a qualified competitor in the XPRIZE Autonomous Wildfire Response competition (advanced to the finalist-announcement stage).
Reporting to the Head of Software: scoped and executed cross-functional initiatives across the AV stack, pairing deep research into individual components with integrated solution design — and embedding into critical-path efforts when the schedule demanded it.
Led a four-engineer team building medium- and long-term mapping and localization technology for industrial autonomous mobile robots; key contributor to AMR-division strategy, requirements, system architecture, and evaluation.
Technical lead of a four-person localization team: multi-sensor SLAM, confinement-safety logic, and algorithm regression testing for the Terra robotic lawnmower. Principal author of the Python analytics framework used for post-mission log analysis and algorithm evaluation.
Built a 1,500-lb AGV that lifts and transports cars: two custom PCBs, STM32 firmware, and ROS-based control software — including techniques to park 7×14-ft steel trays to within inches.
Led a four-person team advancing current and next-generation wrist-wearable heart-rate algorithms; drove product and release strategy, ran longer-horizon exploratory work, and delivered algorithms, circuits, tooling, and IP that other teams productized.
Multi-year engagement: production C++ prediction library, full nRF52 BLE firmware with bootloader/OTA, PDE-based thermal modeling, Kalman instant-read estimation, and manufacturing programming tools. Shipped to consumers.
STM32 firmware, a RAW/Bayer imaging pipeline with color science, and a full AWS IoT backend (device management, OTA, telemetry, factory provisioning) cost-engineered from 500 to 50,000 devices.
Designed the electronics and on-board software (with software-in-the-loop verification) for a fleet of robotic floating sensors studying San-Joaquin delta hydrodynamics; built a robust collision-avoidance and path-selection controller; helped assemble 140 robots and run field studies.
Demos & notes
Interactive walkthroughs of the systems work — safety-critical autonomy, edge-AI pipelines, and the math behind them.
The fleet-ops dashboard from my XPRIZE Wildfire project at Think Circuits, embedded live: detection and suppression drones over the demo site, a scripted incident timeline, wind-adjusted dispatch, and a coverage map that re-skews with the wind. The real React/Mapbox app — only the MQTT backbone is swapped for an in-browser scenario player.
Browser-playable HJB reachability: a warehouse AMR avoiding people, a quadrotor pursuing under wind, and a six-site interceptor network. Built on a CUDA solver; runs at 60 fps on a precomputed value function.
A DSL for safety-critical reflex controllers — worst-case execution time, bounded memory, and safety invariants proven at compile time. Compiles to MISRA-C and synthesizable Verilog from the same source. The compiler runs in your browser; live IDE embedded in the post.
More coming: edge-AI pipelines on Jetson, hardware build logs, and field-test footage.
Experience
Education
Awards & honors
What I work with
ROS / ROS2 autonomy stacks · multi-sensor SLAM & localization · CNN perception · VLM workloads · sensor fusion (Kalman / particle filters) · AprilTag visual fiducials · real-time control · teleoperation (WebRTC)
On-device inference · model deployment to DSP / NPU (Qualcomm Hexagon, HVX SIMD) · NVIDIA Jetson · STFT/iSTFT + CNN/LSTM audio models · wearable biosignal algorithms · post-deployment log analytics
Confinement / functional safety · WCET analysis · regression & verification frameworks · safety-critical DSL design · AI-assisted code validation · FDA & BLE-SIG regulatory pathways
C / C++ / Python · Nordic nRF5x · STM32 · ARM Cortex-M · RTOS (QuRTOS) · BLE incl. bootloader/OTA · PCB & sensor integration · AWS IoT at 50k+ devices · PDE thermal modeling · color science / imaging pipelines
Patents & publications
Patents from Think Circuits, Fitbit, and others; peer-reviewed work from a Berkeley Ph.D. in estimation and field robotics — the research highlights, explained → · Google Scholar profile →
Plus conference papers (ICRA, CASE, DCoSS, ICNP, UbiComp) and a decade of invited talks. Full list available on request.
Let's talk
Open to conversations with frontier AI labs, robotics and hardware product companies, and autonomy teams working on embodied agents and edge deployment.