The Robots Are Coming – Physical AI and the Edge Opportunity

By Pete Bernard
CEO, EDGE AI FOUNDATION

We have imagined “robots” for thousands of years, dating back to3000 B.C. when Egyptian water clocks used human figurines to strike hour bells. They have infused our cultural future with movies like Metropolis  in 1927 through C3PO and R2D2 in Star Wars and more.

Practically speaking, today’s working robots are much less glamorous. They have been developed over the past decades to handle dangerous and repetitive tasks and resemble nothing like humans.  They roll through warehouses, mines, and deposit fertilizer on our farms. They also extend our perceptual reach through aerial and ground-based inspection systems, using visual and other sensor input.

Now that edge AI technology has evolved and getting ever more mature, the notion of physical AI is taking hold and it promises to be a critical platform that is fundamentally enabled by edge AI technologies. A generally agreed definition of physical AI is:

A combination of AI workloads running on autonomous robotic systems that include physical actuators.

This is truly “AI in the real world” in that these systems physically interact with the real world through motion, touch, vision, and physical control mechanisms including grasping, carrying and more. It can combine a full suite of edge AI technologies in a single machine. Executing AI workloads where the data is created will be critical for the low latency and low needs of these platforms. These could range from:

• tinyML workloads running in its sensor networks and sensor fusion
• Neuromorphic computing for high performance/ultra-low power, fast latency and wide dynamic range scenarios
• CNN/RNN/DNN models running AI vision on image feeds, LIDAR or other “seeing” and “perceiving” platforms
• Transformer-based generative AI models (including reasoning) performing context, understanding and human-machine interface functions

These are designed all into one system, with the complex orchestration, safety/security and controls needed for enterprise grade deployment, management and servicing. In addition, as the TOPS/watt and lower power/higher performance edge AI platforms come to the market, this will positively impact the mobility, cost and battery life of these systems.

Robotics is where AI meets physics. They require sophisticated physical capabilities to move grasp, extend sense and perform a wide range of tasks, but they are also software platforms that require training and decision making, making them prime candidates for one of the most sophisticated combinations of AI capabilities. The advent of accelerated semiconductor platforms, advanced sensor networks, sophisticated middleware for orchestration, tuned AI models, emerging powerful SLMs, applications and high-performance communication networks are ushering in a new era of physical AI.

Let’s level set with a taxonomy of robots and a definition of terms. There are many ways to describe robots – they can be sliced by environment (warehouse) or by function (payload) or even by mobility (un-manned aerial vehicles). Here is a sample of some types of robots in deployment today:

• Pre-programmed robots
  • These can be Heavy Industrial robots, used in very controlled environments for repetitive and precise manufacturing tasks. These robots are typically fixed behind protective barriers, costs hundreds of thousands of dollars.

• Tele-operated robots
  • These are used as “range extenders” for humans to perform inspections, observations, or repairs in challenging human environments – including drones or underwater robots for welding and repair. Perhaps the best-known tele-operated robots were the robots sent to Mars by NASA in the last few decades. There has also been a fish robot named SoFi designed to mimic propulsion via its tail and twin fins, swimming in the Pacific Ocean at depths of up to 18 meters. [1]

• Autonomous robots
  • You probably have one of these in your house in the form a vacuum cleaner robot navigating without supervision and relying on its sensors for navigation. Recently we have seen a number of “lawnmower” robots introduced to take on this laborious task. In Agriculture, robots are already inspecting and even harvesting crops in an industry with chronic labor shortages[2]. There is also a thriving industry for autonomous warehouse robots – including in Amazon warehouses. [3]

• Augmenting robots
  • These are designed to aid or enhance human capabilities such as prosthetic limbs or exoskeletons. You probably first were exposed to this category of robots when you watched The Six Million Dollar Man” on TV –but on a more serious note, they are providing incredible capabilities for amputees and enabling safer work environments for physical labor.[4]

• Humanoid robots
  • Here’s where it gets interesting. We have developed a bi-pedal world – why not develop robots that work in that world as it’s been designed? Humanoid robots represent humans – as bi-pedal (or quad pedal in the case of Boston Dynamics), can communicate in natural language and facial expressions and perform a broad range of tasks using their limbs, hands and human-like appendages. The number of quad-pedal robot have only been deployed in the low thousands worldwide and we are still in the very early stages of development, deployment, and reasonable cost. Companies like Enchanted Tools[5] are demonstrating humanoid robots that can move amongst humans for carry lighter loads, deliver items, and communicate in natural language. Although humanoid robots will catch the bulk of the attention of the media in coming years, and face the most “cultural impact,” the other robot categories will also benefit greatly from generative AI and drive significantly greater efficiencies across industries.

How Generative AI on the edge will impact Physical AI

It’s hard to overstate the impact that Generative AI will have on the field of robotics. Beyond the ability for much more natural communication and understanding, Generative AI model architectures like Transformers will be combined with other model architectures like CNNs, Isolated Forests and others to provide context and human machine interfaces for image recognition, anomaly detection and observational learning. It will be a “full stack” of edge AI from metal to cloud.

Let’s take a look at the differences between traditional AI used in robotics and what Generative AI can bring:

In summary, while many forms edge AI are excellent and necessary for analyzing existing data and making predictions in resource constrained and low power environments, generative AI at the edge will now add the ability to create new data and adapt dynamically based on experience. The application of Generative AI to robotics will unlock observational learning, rich communication,  and a much broader application of robots across our industries and our lives.

Safe and Ethical Robotics

Whenever robots are mentioned, the comparison to
“evil robots’ from our culture are not far behind. The Terminator, Ultron or Gunslinger from Westworld. And at the same time, we have enjoyed anthropomorphized robots like C3PO and R2D2, or Wall-E. And then there are ones in -between, like from the movie The Creator.

As attention has been paid to the scope Generative AI moving to AGI, what guardrails, best practices and outright legislation exists to keep robotic efforts – pared with Generative AI – in the category of good or neutral?

Isaac Asimov famously penned his three laws of robotics back as part of his short story “Runaround” in 1942:[6]

• A robot shall not harm a human, or by inaction allow a human to come to harm
• A robot shall obey any instruction given to it by a human
• A robot shall avoid actions or situations that could cause it to come to harm itself

In 2021, Dr. Kate Darling – a research specialist in human-robot interaction, robot ethics and intellectual property theory and policy at the Massachusetts Institute of Technology (MIT) Media Lab – wrote an article in The Guardian proposing that we think about robots more like animals than a rival to humans. Once we make that shift, we can better discuss who are responsible for robot actions and who is responsible for the societal impacts that robots bring, such as transformations in the labor market.[7]

The European Union published “Civil law rules on robotics” back in 2017 that addressed the definition of a robot, where liability lies, the role of insurance and other key items. In 2023 a law was introduced in Massachusetts in the US that would 1) ban the sale and use of weapons-mounted robotic devices, 2) ban the use of robotic devices to threaten or harass, and 3) ban the usage of robotic devices to physically restrain an individual. It’s unclear how or when similar legislation will make it to the federal level.

Observational Learning Is a Game Changer

In the world of edge AI, training has happened on “the cloud” or in server-class GPU environments and inferencing has happened on the light edge. With the introduction of reinforcement learning and new work in continuous learning we will see the edge becoming a much more viable area for training.

However, in physical AI platforms, observational learning (sometimes referred to as behavior cloning) in AI allows robots to learn new skills simply by watching humans – in reality or in a simulated physical environment. Instead of being programmed step-by-step, robots can make connections in their neural networks based on observing human behavior and actions. This kind of unstructured training will enable robots to better understand the nuances of a given task and make their interaction with humans much more natural.

There have been a number of key advanced in AI models for observational learning, starting with CNN model types and recently leveraging diffusion model types such as the one presented in the Microsoft Research paper in 2023 – Imitating Human Behaviour with Diffusion Models.[8]

In March of 2024, NVIDIA introduced Gr00t[9], their own foundational model designed for observational learning of their ISAAC/JETSON robotics platforms. It was demonstrated at the NVIDIA GTC keynote by Jensen Huang and also leverages their Omniverse “digital twin” environment to develop virtualized physical environments that can train robots via observational learning in a safe and flexible virtualized environment. This was updated in 2025 to Gr00t N1 as well as a new “Newton” physics engine. We’re now seeing Foundation models tuned for robotics platforms[10] like Gr00t, but also RFM-1 by Covoariant, among others. Expect this area to proliferate with options much like Foundation models for LLMs in the cloud.

Robotics as a “three computer problem” – there is an AI model training in the cloud using generative AI and LLMs, there is model execution and ROS running on a robotics platform itself, and a simulation/digital twin environment to safely and efficiently develop and train.

The Edge AI Opportunity for Robotics 

 “Everything That Moves Will Be Robotic” – Jensen Huang

The confluence of generative AI and robotics is swinging the robotic pendulum back into the spotlight. Although Boston Dynamics has only deployed around 1500 Spot robots worldwide so far, expect many more, and in many more configurations, throughout our warehouses, our farms, or manufacturing floor. Expect many more humanoid experiments and expect a hype wave washing over us with plenty of media coverage of every failure.

Running generative AI on these platforms will require significant TOPS horsepower as well as high performance memory subsystems in addition to advanced controls actuators and sensors. We will see “datacenter” class semiconductors moving down into these platforms but just as interesting will be edge native semiconductor platforms moving up into this space, with the kinds of ruggedized thermal and physical properties as well as low power and the integrated communications needed. We will also see many new stand-alone AI acceleration silicon paired with traditional server class silicon. Mainstream platforms like phones and AI PCs will help drive down costs with their market scale.

However, in addition to requiring top end semiconductors and plenty of RAM, robotic platforms – especially humanoid ones – will require very sophisticated sensors, actuators, and electro-mechanical equipment – costing tens of thousands of dollars for the foreseeable future.

To keep things in perspective, Goldman Sachs[11] forecasted a 2035 Humanoid Robot TAM of US$38bn with shipments reaching 1.4m units. That’s not a tremendous unit volume for humanoid robots (PCs ship around 250m units per year, smartphones north of a billion) – we can expect orders of magnitude more “functional form factor robots” in warehouse, vacuuming homes and doing other focused tasks.

These platforms – like the ones now available from Qualcomm, NVIDIA, NXP, Analog Devices and more – are attracting developers that are taking their server class software skills and combining them with embedded computing expertise. Like mobility, robotics and physical AI are challenging developers and designers in new ways and provides a unique opportunity for workforce development, skill enhancement and career growth.

A key challenge here is to avoid the pitfalls of Industry 4.0 and IoT – how do we collaborate as an industry to help standardize on data sharing models, digital twin models, code portability and other elements of the robotics stack? If this area becomes more fractured and siloed we could see significant delays in real deployments of more advanced genAI driven robots.

Developers, designers and scientists are pushing the envelope and closing the gap between our imaginations and reality. Like with cloud-based AI, the use of physical AI will require important guardrails and best practices to keep us not only safe but make this newfound expansion of physical AI capabilities accretive to our society, but the future

We cannot underestimate the impact that new robotics platforms will have on our culture, our labor force, and our existential mindset. We’re at a turning point as edge AI technologies like physical AI are leveraging traditional sensor AI and machine learning with generative AI, providing a call-to-action for all technology providers in the edge AI “stack,” from metal to cloud, as well an opportunity for business across segments to rethink how these new platforms will leverage this new edge AI technology in ways that are still in our imagination.

[1] https://www.csail.mit.edu/research/sofi-soft-robotic-fish

[2] https://builtin.com/robotics/farming-agricultural-robots

[3] https://www.aboutamazon.com/news/operations/amazon-introduces-new-robotics-solutions

[4] https://www.automate.org/robotics/service-robots/service-robots-exoskeleton

[5] https://enchanted.tools/

[6] https://www.goodreads.com/en/book/show/48928553

[7] https://tdwi.org/articles/2021/06/16/adv-all-building-ethical-guardrails-into-ai-driven-robotic-assistants.aspx

[8] https://www.microsoft.com/en-us/research/publication/imitating-human-behaviour-with-diffusion-models/

[9] https://nvidianews.nvidia.com/news/foundation-model-isaac-robotics-platform

[10] Foundation Models in Robotics: Applications, Challenges, and the Future – https://arxiv.org/html/2312.07843v1

[11] https://www.goldmansachs.com/intelligence/pages/gs-research/global-automation-humanoid-robot-the-ai-accelerant/report.pdf