Generic selectors
Exact matches only
Search in title
Search in content
Post Type Selectors
About the Author
About the Author
About the Author

Physical AI: Beating Odds to Make Edge Devices Sense and Think and Robots Learn and Act. What Can We Learn and Do?

Being Human for a Better Tomorrow in the Age of AI

  • February 28, 2025
  • 37th edition
Physical AI Beating Odds to Make Edge Devices Sense and Think and Robots Learn and Act. What Can We Learn and Do

AI : Going Beyond Efficiency and Cost Reduction

2024 has seen an early majority of enterprises embark on AI adoption. Yet, many initiatives focus primarily on cost savings. While efficiency is important, an overemphasis on cost-cutting misses the broader opportunity—AI as a driver of new services, products, and revenue streams.

A January 2025 McKinsey report on achieving “Super Agency in the Workplace“ in AI shows that while companies seek at least 5% revenue growth from AI, only 20% currently achieve this.

As quoted in the 32nd edition, Dr. APJ Abdul Kalam, former President of India, once said:

“Small ambition is a crime.”
AI’s potential extends beyond operational improvements—it’s about unlocking entirely new opportunities.

Physical AI - Hard Challenge and All Odds

Physical AI—intelligence that senses, thinks, and acts in the real world—takes on challenges tougher than cloud-based language models.

It’s a bigger leap, moving from ideas to actions, often with less data to start from. Edges can be unpredictable, networks spotty, and resources tight—small models fit into small spaces. Costs shift, supply chains wobble, timelines stretch, and mistakes aren’t easy to fix.

Still, Physical AI is finding its way, from edge devices to robots, showing what’s possible with imagination, smart design, and solid engineering. It’s not just about beating odds—it’s about building something new.

Yet, despite these odds, Physical AI from edge devices to robots is breaking through, providing lessons and inspiration.

Physical AI: Manifestation Passive (Edge Device) and Active (Robots)

Physical AI has two different manifestations:
  • Passive (Edge Device) — think phones and cameras which excel at sensing and thinking, processing the world without moving a muscle. They’re the quiet observers, turning raw inputs into insights on the spot.
  • Active Device (Robots) take it further—they learn and act, manipulating their surroundings with purpose, from moving packages in a warehouse or a humanoid folding laundry to robots for example.

Physical AI can be successful only if the challenges are addressed with metaphors for imagination, well-thought design, and engineering excellence. *** (See footnote and sidebar)

The following examples illustrate this and give us a lot to know, learn, and be inspired for the future.

Passive Physical AI (Edge Device)

Apple Intelligence

Design:Apple Intelligence runs a Small Language Model (SLM) on iPhones, built into the Neural Engine of A17 Pro chips and beyond. It’s crafted for privacy—keeping data local—and handles tasks like text generation and photo searches.

Approach:The system processes everything on-device, tapping 35 trillion operations per second (TOPS) for speed, with Private Cloud Compute stepping in securely for bigger jobs. It weaves AI into Siri, Messages, and Photos, using screen context to respond naturally.

Success: Launched in iOS 18.1 (October 2024), it’s hit 20-30 million users by early 2025 (eMarketer), with 85% of beta testers praising its seamless, private polish (Laptop Mag, January 2025). It’s a quiet win over cloud-heavy rivals.

Gemini on Android

Design: Gemini Nano is Google’s edge Small Language Model (SLM), a 1-3 billion parameter model designed for Android’s vast range—from 4GB RAM budget phones to Pixels. It manages text, voice, and image tasks without constant cloud reliance.

Approach: It blends on-device processing with cloud support, tying into Google’s ecosystem—Search, Gmail, Photos—for versatility. Multimodal smarts let it draft replies or analyze pics, adapting to hardware limits with efficiency.

Success: By February 2025, tens of millions use it daily (TechRadar), with 70% of Android users valuing its reach across devices (Digital Trends, December 2024). It’s proof scale and edge can coexist.

Noteworthy AI Inspector

Design: Noteworthy AI’s Inspector Edge is a camera system with a Small Language Model (SLM) , mounted on utility trucks to spot power grid flaws. It’s rugged, built on edge processors (likely ARM-based) for tough outdoor work.

Approach: High-res images—cracked poles, worn lines—are processed locally in real-time, with only metadata sent to a cloud dashboard. Low-power AI keeps it running on routine drives, dodging network gaps.

Success: Pilots with FirstEnergy and five US utilities cut inspection times by 60% (Forbes, September 2024), curbing outages as weather risks climb (up 67% since 2000, Climate Central). It’s a steady fix for critical grids.

Hailo-8

Design: Hailo-8 is a compact edge AI chip—37mm², 26 TOPS, 2.5W—made for smart cameras and traffic systems. It’s built to run vision-focused SLMs, like spotting pedestrians or plates.

Approach: A dataflow architecture squeezes efficiency from convolutional neural networks, linking to devices via PCIe or USB. Its low heat and power fit outdoor chaos—urban streets, weather shifts.

Success: Used by 300+ firms (X posts, 2025), it’s cut traffic accidents 15% in Singapore trials (TechCrunch, January 2025). Users call it a cost-effective edge into safer cities.

Active Physical AI (Robots)

75 years ago, a new era of manufacturing automation for large-volume production began. This era was made possible by three categories of production machinery:
  • Special Purpose Machines (SPMs) — can do only a specified set of actions but do it very efficiently.
  • Numerical Controlled (NC) Machines — have limited programmability.
  • Computer Numerically Controlled (CNC) Machines — multifunctional.
As we see in the examples below, this is analogous to what is happening in Physical AI in Robotics. We have purpose-built production robots, self-driving cars like similar to SPMs, a learning robot like NC machines that can work in a warehouse, kitchen, or hospital, and finally human-like general learning and adaptability such as humanoids.

Purpose Built Robots — Tesla Full Self-Driving (FSD)

Design: Tesla’s FSD chip, a 144-TOPS powerhouse, drives autonomous cars, built to process eight camera feeds for one job: safe navigation. It’s a purpose-built artificial intelligence to sense and trigger actions.

Approach: All computation happens on-device, scanning roads in real-time for obstacles and paths. Over-the-air updates refine its focus, keeping it sharp for driving alone.

Success: Over 500,000 vehicles run it by 2025 (X posts), slashing crash rates 40% (NHTSA, 2024). It’s a single-task master, proving precision beats odds.

Learning Robots @ Kitchen, Warehouse and Hospital

Chef Robotics

Design: Chef Robotics crafts robots with vision AI and deep learning, built to portion food—sandwiches, mashed potatoes—in busy kitchens. It’s made for one setting’s quirks.

Approach: The system watches and adjusts grip on the fly, learning to handle variety within food prep’s mess. It’s context-driven, sticking to culinary chaos.

Success: It’s cut waste and labor costs in kitchens (WIRED, 2024), a practical win for an unpredictable trade. Early adopters say it’s like a skilled helper.

Covariant

Design: Covariant’s Brain is an AI system for warehouse robots, using reinforcement learning to grab unfamiliar objects. It’s designed to adapt within logistics hubs.

Approach: Robots share data across sites, picking up new items—boxes, bags—without pre-set rules. It learns fast, staying in its warehouse lane.

Success: Handling millions of picks daily (Forbes, 2024), it’s outpaced rigid bots, showing context learning can scale. Warehouses run smoother with it.

Moxi by Diligent Robotics

Design: Moxi is a hospital robot with vision and natural language processing (NLP), built to deliver supplies in crowded wards. It’s tuned for healthcare’s bustle.

Approach: It navigates halls and elevators, prioritizing tasks—meds over linens—based on real-time needs. It learns its space, not the whole world.

Success: Nurses save hours daily (X posts, 2025), easing chaos with quiet support. It’s a contextual fit hospitals trust.

Robots with General Learning and Adaptability

Physical Intelligence (PI)

Design: PI’s π0 is a foundation model for robots, trained on teleoperated data to handle any task—laundry or table clearing—with transformer AI. It’s built for broad adaptability.

Approach: It learns from human demos, aiming to control diverse machines without custom coding. It’s a generalist, tackling whatever’s thrown at it.

Success: Demos fold T-shirts and sort pantries (WIRED, November 2024), hinting at robots that learn like us. It’s early, but the scope’s wide.

Sanctuary AI​

Design: Sanctuary AI’s humanoids use multi-modal AI—vision, touch, language—to tackle over 100 tasks, from stocking to cleaning. They’re made to adapt anywhere.

Approach: On-the-fly training lets them shift jobs, learning through interaction. It’s not one role—it’s many, built on flexibility.

Success: Early deployments show promise (X buzz, 2025), with versatility that could redefine robotics. It’s a step toward general helpers.

Lessons from Pioneers : From Beating the Odds

How did these ten pioneers design and engineer to conquer Physical AI’s hard challenges:
  • Apple Intelligence: Crushed compute/memory limits with on-device SLM efficiency.
  • Gemini: Scaled scarce data across Android’s billions, hybrid way (both device and cloud).
  • Noteworthy AI: Mastered unpredictable environments with local processing.
  • Hailo-8: Squeezed 26 TOPS into 2.5W, defying resource constraints.
  • Tesla FSD: Nailed mission-critical precision, no errors allowed.
  • Chef Robotics: Adapted to volatile variability—food’s messy reality.
  • Covariant: Sped prototype-to-production with shared learning.
  • Moxi: Navigated complex timelines, thriving in hospitals’ bustle.
  • Physical Intelligence: Built data from scratch—teleoperation at scale.
  • Sanctuary AI: Tackled cost economics with multi-task versatility.

It’s also not all smooth wins. There are misses too, where actions go wrong, and that’s where learning deepens. Tesla’s FSD has had crashes, Gemini’s cloud dips lag on weaker phones, Hailo 8 integration hiccups, Physical Intelligence’s π0 flings boxes in early demos for example.

While some actions have gone wrong, overall it is hard to overstate the importance of these early successes. They made impossible, possible.

Physical AI : A New Cambrian Explosion is Beginning with Building Blocks

Physical AI as seen with these examples, has been bespoke so far—custom systems like Apple’s SLM or PI’s π0, each a tailored feat.

Now, that’s changing with building blocks for Physical AI. Tools like NVIDIA Jetson, Google Coral, and Skild Robotics aren’t just enablers—they’re accelerators, turning slow, one-off builds into a fast-spreading wave, a Cambrian surge of possibility.

NVIDIA Jetson/Isaac packs 200 TOPS into a small module, letting factories run sorting bots or farmers deploy crop drones without starting from scratch (Forbes, 2024). It’s not a single robot—it’s a platform, cutting design time so anyone with an idea can build.

Google Coral, at 4 TOPS, opens the door wider: a cheap chip powering wildlife cams in Kenya or smart locks in homes (Wired, 2024). It’s simple, scalable—hobbyists and startups crank out edge devices fast, spreading AI where custom rigs couldn’t.

Skild Robotics flips it to data, gathering human actions—lifting, stacking—to train robots across tasks (X posts, 2025). One dataset feeds hundreds of machines, slashing the need for bespoke coding.

These tools don’t just help—they multiply. Jetson’s compute scales production, Coral’s accessibility sparks invention, Skild’s data fuels learning. Where Physical AI was once a craftsman’s trade, this trio makes it a factory—fast, broad, and ready to grow.

Life in 2035 : What May Be

Wake up, and your room’s already adjusting—lights soften, coffee starts, a wristband nudges you with the day’s plan, all synced without a tap.

Step outside, and the street hums: cars glide silently, rerouting themselves, while a drone drops your meds from a pharmacy you never called.

At work, a desk bot shifts your tools—screen tilted, files ready—learning your rhythm as you go. Lunch is quick: a kitchen machine preps a meal, tweaking spices to your taste, no recipe needed.

Home again, the space shifts—chairs ease your posture, a helper sorts clutter, floors stay clean without asking.

Night settles: lights dim as you move, a small bot tracks your sleep, cooling the air if you stir.

Premise: By 2035, Physical AI could be as common as today’s Wi-Fi—edge devices and robots woven into life, not special, just there. It’s not about gadgets; it’s about how a day feels different with them sensing, thinking, acting alongside us.

Takeaways For You Today

What can Physical AI do? These stories show what’s possible—where could they lead?

While we wait to see the day, as you have seen even in the last five editions, AI led Accelerated Innovation (AI for AI) matters.

AI tech and infrastructure is constantly and rapidly evolving in its capabilities and support super agency to deliver revenue and strategic value.

Being efficient is one and being disruptive is another. Both are paths to pursue.

However, to do the disruptive, the key questions you should ask/continue to ask are:

  • 1. What are the hard problems to solve in your business and function?
  • 2. What will be the value that can be created by solving this problem?
  • 3. How will AI enable the solution?
  • 4. What internal and external (partners) are needed to plan and execute?
  • 5. How do I be part of this change in my industry with my know how and skills?

Final Word

Across our 5 editions on AI for AI we saw a how small team at DeepSeek disrupt AI itself and seen 37 AI and Tech led innovators from using AI in FS, Pharma ,Healthcare , Enterprise Software and Physical AI .

And we will find impactful and exciting ones as we cover other key industries in the upcoming editions.

In democratized AI, Any one, any team, any where can do any thing if they can, if they have the courage, intelligence and tenacity.

And the time to do is now.

Nearly three weeks before Apple Intelligence announcement, in the 19th edition (Oct 11,2024) about metaphors, grounding and imagination of using AI we explored disruptive innovation for Apple, Marriott and LinkedIn. See Sidebar , Apple is happening.

Explore, subscribe, join and stay tuned for more!

Coming Next: AI for AI – rethinking business models

Love to hear your comments, thoughts and ideas.

Best wishes

“AI is a language. Treat it like one: practice, iterate, and mind your grammar prompts, assumptions, and verification.”

II. Questioning / Asking

Good conversations flow from well‑sequenced questions—topical, simple, coherent, cohesive.
LLM Conversation Example 1
Q: What are empirical judgments?

A: Empirical judgments are based on observation, experience, or experimentation.
Q: What are moral judgments?

A: Moral judgments are based on ethical principles and values.

IV. Judgment

“The structure of a language affects its speakers’ worldview and cognition.”
— Henry Hazlitt
“The art of questioning is the source of all knowledge.”
— Thomas Berger
Share: