Physical AI Meetup

Hear 3 expert talks, take the mic for an AI/hardware demo or announcement, and connect over pizza & drinks.

Important: Please register via Eventbrite for your [free ticket here] – registration is required and spots are limited.

A big thank you to Innovation Central Perth (ICP) at Curtin University for hosting this meetup.

Take the mic at our community open mic. Share a quick demo, prototype, or announcement – whether it's AI, hardware, or something you're building. All welcome.

Agenda:

**Pizza & Networking**
Enjoy pizza and drinks while connecting with other attendees.

**Welcome & Intro**
Hear a short welcome from one of the meetup organisers.

**Massimo Sacchi: Democratisation of Edge AI**
Overview: This talk brings Physical AI down from theory to the factory floor. It opens by framing Edge AI along two axes—edge versus cloud, and learned models versus traditional analytical rules—then walks through why industrial environments increasingly demand on-device intelligence. Success stories anchor the talk on real Arduino hardware, covering everything from the first POC to product industrialization, with minimal re-engineering effort. The talk closes by looking toward continuous learning at the edge as the next frontier for industrial AI.

Speaker bio: Massimo is an industrial automation expert with more than 25 years of experience in multinational contexts. His passion for innovation and new technologies drives his interest in IIoT, cloud applications, and AI, as well as the implementation strategies of Industry 4.0 concepts. He started his career as an Automation Engineer in heavy industry, later worked as Automation Manager in logistics and as Product Manager for a worldwide leading manufacturer of industrial sensors. Since moving to Australia in 2015, and from 2019 as a consultant, he has supported customers worldwide, including Arduino, where he was Business Developer and Corporate Partnerships Manager, helping expand the company in the industrial solutions market.

**Stanley Chong: 20ms vs 2 Seconds: What a Safety Vest Taught Us About Choosing the Right Visual AI**
Overview: Not every vision problem needs a giant model — and not every giant model needs to see pixels the way you'd think. In this talk we'll unpack the real differences between CNNs, Vision-Language Models (VLMs), and multimodal LLMs: what they're each actually built to do, where their reasoning comes from, and why picking the wrong one costs you either accuracy or your latency budget. Then we'll put it to the test live, on a Jetson running our own safety-vest compliance demo (built with our ICP intern). We'll run the same camera feed through a YOLO object detector (20ms) and a PaliGemma VLM (2 seconds) side by side, and watch where each one shines — and where each one quietly gets it wrong. If you're building physical AI and wondering whether you need a 2-billion-parameter model or a 20-millisecond one, come find out.

Speaker bio: Stanley has spent the last 30 years in motion controls, distributed control systems, and industrial IoT. These days he splits his time between industrial monitoring systems work and a part-time role as Technical Officer at ICP, where he supports student interns tackling both internal and industrial projects. His current curiosity sits at the edge: how edge computing, industrial IoT, and cloud-based ML/AI actually fit together in the field, and what breaks when you try to run a vision model on a fraction of a cloud GPU's power.

**Dr Fangyi Zhang: Gentle Embodied AI: Vision- and Tactile-Based Robotic Manipulation for a Delicate World**
Overview: Modern robots excel at tasks demanding force, speed, and repeatability, yet they remain remarkably poor at gentleness – the careful, adaptive, contact-rich interaction required to handle fragile objects, manipulate deformable materials, or physically collaborate with people safely. This talk presents the research program of Dr Zhang's group, which centres on gentle and dexterous robotic manipulation. It begins with vision-based manipulation and learning-based control – how robots can acquire visuo-motor skills, transfer them from simulation to the real world, and learn manipulation behaviours by watching human demonstrations. It then turns to gentleness and compliance: sensing and reasoning about contact and force so a robot can modulate its grip and motion, which is what gentle handling actually requires. A central theme is bringing vision, learning, and contact information together so robots can adapt to delicate, deformable, and unpredictable objects – one of the open challenges in contact-rich manipulation. The talk shares representative results, lessons learned along the way, and the broader directions that look most promising for the field and for collaboration.

Speaker bio: Dr Fangyi Zhang is a Lecturer at Curtin University, where he is building the Embodied Robotics and AI (ERA) Lab. His research centres on gentle and dexterous robotic manipulation, combining vision-based perception, robot learning, and tactile and force sensing to help robots interact with the physical world safely and carefully. He has 23 peer-reviewed publications across leading robotics and machine learning venues including IJRR, RA-L, ICRA, IROS, NeurIPS, and ICLR. Before joining Curtin, he worked with the Australian Centre for Robotic Vision, QUT Centre for Robotics, and Alibaba DAMO Academy. Since 2023, he has served as an Associate Editor for IEEE Robotics and Automation Letters.

**Dr. Joe Peng: Closing Keynote From (A)mbition to (I)nnovation**

**15-minute Panel Q&A**

**Community Open Mic**
Short informal shares – demo, prototype, or quick announcement.

**Wrap-Up**
Any further Q&As and chats will happen here.

Last reminder: A free Eventbrite registration is required to attend – [register here]