By Giorgio Ippoliti – technologist, field applications engineering EMEA, Sandisk.
While Artificial intelligence (AI) has been particularly good at answering questions, many may be wondering how else the technology could help us in our daily lives. Well, we are already starting to see the next phase of AI, where intelligence has leapt off the screen and into the physical world.
This transition, known as ‘Physical AI’, enables machines to perceive, understand, reason, and have natural interactions with the world. As this shift accelerates, it is shaping how intelligent systems are built, deployed, and scaled, with key implications for edge computing, latency, and high-performance local data storage.
We are already living amidst physical AI, for example autonomous vehicles, drones and, of course, robots. These are technology marvels, but the next transition in physical AI, the humanoid robotic era, is even more complex.
From Collaborative Edge AI to the Next Frontier in Robotics
The early 1970s marked the beginning of mobile robots3, which mainly specialized in welding and material handling. These real-time operating systems (RTOS)-based robots ensured smooth operation where precise scheduling was required at scale and were controlled centrally. These had limited storage locally, only a few kilobytes (KB) of RAM, and data was sent and kept on external computers or PCs.
Since 2019, AI has been evolving rapidly and has extended out from a central cloud architecture to operate at the edge, for real-time, localised inference and decision-making. We call this the ‘collaborative edge AI’ and it is critical in situations where milliseconds matter, for example in autonomous vehicles.
Another example is an autonomous mobile robot (AMR), which is equipped with autonomous path planning and dynamic obstacle avoidance intelligence, say, within warehouse environments. As on-edge intelligence or inference increases, so does the need for edge or local storage. Moreover, as these AI devices need faster access to data, the need for high-performance flash storage becomes central for their smooth operation. Today, these devices incorporate flash storage in gigabytes (GB).
What’s coming up could be one of the most exciting eras in AI, the humanoid robot era.
A New Era of Humanoid Robots Begins
Humanoid robots are projected to reach 1.4 million units by mid 2030s and the figure may be revised upward, as investment in the sector is growing faster than anticipated1. This physical AI use case relies on fast, NVMe powered flash storage. As, in addition to training data residing locally, physical AI will infer and learn continuously, without the need to code everything by hand, which will further inflate the amount of on- device data and therefore the need for local storage.
For example, from learning how much pressure a soft fruit can withstand and what happens when that limit is crossed, to understanding what happens when a glass is not properly balanced on a table, physical AI learns continuously from interaction with the surrounding world, building on the original data it holds. Also, real-time recordings from multiple cameras, within these robots, for anomaly detection and simulation replay will lend to more storage requirements.
The storage on the edge is being used for decision making and caching, and important data is sent back to the cloud for long term storage and centralised model training.
Solving the Storage Paradox
However, it’s not a one-size fits all approach, as different humanoid robot types will have different storage requirements. An industrial humanoid robot stores large Vision Language Action (VLA) datasets, multi-cameras light detection and ranging (LiDAR), local AI inference models, high-frequency motion logs, and maintenance telemetry. It may need multiple terabytes (TB) of NVMe solid state drives (SSDs).
Whereas a general purpose ‘embodied humanoid robot’ will run multimodal foundation models and world models locally covering speech, vision, and control, often with large language models (LLM) and diffusion models deployed at the edge to reduce latency. These may need twice as much local storage as an industrial humanoid robot, while field or hazard-zone humanoid robots could demand even more storage than a general-purpose humanoid robot locally or on edge. Sandisk offers a broad portfolio of m.2 NVMe SSDs and embedded flash products to address these use cases.
A large amount of data collected by humanoid robots will be sent to data centers (DC) for training and inference purposes. This brings a major shift in DC architecture. AI workloads require more frequent and faster access to data, driving the need for faster data lakes. Consequently, high-performance, and high-capacity SSDs based on UltraQLC technology become foundational. Sandisk announced that its SANDISK® 128TB and 256TB NVMe SSDs with UltraQLC technology set a new benchmark for hyperscale flash storage, and are purpose-built for the fast, intelligent data lakes powering AI at scale.
As physical AI becomes the latest sector to develop at lightning speed, the growth and success of humanoid robots will inevitably depend on a foundation built on flash storage.
