ARTIFICIAL INTELLIGENCE What challenges does AI
face in the physical world? By Dr. Massimiliano “Max” Versace vice president, Emergent AI
I
n the physical world, AI needs to face physical signals (not words or pixels, as it has mostly done as of today). In the physical world, there is no luxury for labelled data or handcrafted reinforcements. Also, the physical world is much more uncertain and ‘clean’ than hand-made datasets. AI today works well in software, but it struggles in reality, because the real world is hard. The main problems are as follows: Latency: In the physical world, decisions cannot wait. For instance, a robotic hand cannot pause to think when an object slips. Control loops run in milliseconds, sometimes faster. Systems that depend on remote or centralised compute simply break that loop. Power: Many systems (e.g., autonomous systems) live under strict energy budgets. By adding AI compute, the system becomes smarter, but less useful as its battery depletes. Bandwidth: Sensors produce enormous amounts of data (e.g., vision, tactile, audio). Moving that data to a central processor creates bottlenecks and introduces fragility. Wires and links become points of failure. Variability: Any real-world application faces uncertainty and noisy physical world. Contact
20
is uncertain; lighting changes, materials behave unpredictably. Models trained in clean environments fail in the field. On-device learning and adaptation are a must, and it’s a harder AI set of models to create and field. Orchestration: When you have multiple AI systems, you have to decide where AI runs; what runs locally, what runs centrally, how workloads move across the system and talk to each other. This is still an unsolved problem.
What work needs to be done to conquer these challenges?
First, we must move from centralised AI to distributed AI. Intelligence does not live in one place; it is hierarchical. Reflexes at the edge, reasoning nearby, planning centrally, with each layer solving a different time scale. Second, we must co-design sensing, compute and AI. Traditional systems treat them separately, with sensors collecting data, processors and models interpreting it. This separation is inefficient, creating latency and overhead. Analog Devices (ADI) is aiming to flip this, building systems where sensing and compute are tightly coupled. Third, as a corollary, is the need for new
JULY/AUGUST 2026 | ELECTRONICS FOR ENGINEERS
compute paradigms. Traditional GPUs are powerful but inefficient for edge systems. Emerging approaches, such as in- memory compute, sparse processing, and neuromorphic architectures can deliver orders-of-magnitude gains.
Fourth, we need real-world data loops. E.g., in ADI’s DexAI, tactile intelligence is learned from data that captures real interactions. E.g., force, vibration and motion, which directly encode physical entities, allowing AI to generalise to manipulation tasks.
How will AI improve robotics, drones and autonomous systems?
AI changes these systems in a fundamental way.
Today, many robots execute pre-defined sequences. We are exploring how they can adapt in real time and handle variability. E.g., adjusting forces to correct errors as they happen.
In dexterous manipulation, it is clear that vision alone is not enough. Tactile sensing, our first biological sense, allows systems to understand contact, slip, texture and pressure. This enables stable grasping and precise manipulation and unlocks the real
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46
