AI inference is rapidly expanding from the cloud to the edge, enabling increasingly intelligent experiences across mobile, automotive, industrial, IoT, and embedded systems. This session explores Arm’s AI at the Edge strategy and how technologies such as ExecuTorch are helping translate that strategy into practical on-device AI deployment.

The session begins with a 30-minute in-person presentation from Robert Iannello on Arm’s AI at the Edge strategy, exploring the evolution of AI from cloud-based computing towards increasingly distributed and on-device experiences. It provides an overview of Arm’s approach to enabling AI across a broad range of devices and markets, highlighting the importance of efficient, scalable computing and a strong software ecosystem. It also considers how advances across hardware and software are enabling developers to bring increasingly capable AI experiences to the edge.

Building on this strategic context, a technical presentation by Matt Cossins introduces ExecuTorch and its role in extending the PyTorch ecosystem to edge and embedded devices. The presentation follows the journey from PyTorch model development towards efficient on-device execution, providing an overview of model preparation, ahead-of-time compilation and optimization, runtime deployment, and backend delegation. Together, the two presentations connect the broader evolution of AI at the edge with the practical software workflows developers can use to deploy AI applications across Arm-based devices.

Large Language Models (LLMs) have transformed the field of artificial intelligence. Their extension to computer vision has led to the emergence of Vision-Language Models (VLMs), paving the way for the next generation of Physical AI through Language-Action Models (LAMs) and Vision-Language-Action (VLA) models. These models integrate visual and linguistic information to generate concrete actions for autonomous agents and machines including robotic systems.

VLA models are fundamentally changing robotics. Rather than executing pre-programmed sequences of operations in structured environments, VLA-enabled robots are expected to interpret instructions, reason using external knowledge, adapt to unfamiliar situations, and perform tasks in previously unseen environments. Concurrent advances in world models, agentic AI, reinforcement learning, and foundation models are accelerating the development of intelligent robots capable of autonomous action.

The assumption that VLAs can fully represent the physical world using vision, language, and sensor inputs encoded as tokens is, at present, optimistic. Although these multimodal inputs can be fused through cross-attention mechanisms and transformer architectures can generate action sequences, there remain open research challenges.

This talk will examine the current state of Physical AI, the scientific and engineering challenges that remain, and the opportunities for deploying intelligent machines in complex, dynamic, and unstructured real-world environments.

One promising avenue we pursue in our research group is the development of richer world models that integrate scientific knowledge with data-driven learning. Physics-Informed Neural Networks (PINNs) incorporate physical laws, expressed as differential equations, into machine learning models. However, a comprehensive world model should extend beyond physics and training data and the corpora of general text it was trained on (if the backbone used is an LLM). Knowledge is also encoded in subject-specific textual descriptions, knowledge graphs and other structured representations. To build more robust and generalisable world models, we should exploit all available forms of prior knowledge. We have developed a new framework that enriches parameter discovery of model equations by integrating auxiliary knowledge sources alongside observational data [1, 2].

A second challenge we pursue is the integration of continuous time-series data into VLA models. Industrial robots operate in environments rich in continuous sensor signals, including force, vibration, current, temperature, and motion measurements. However, most LLMs and VLMs are pretrained on discrete modalities such as text and images, making their adaptation to continuous temporal signals non-trivial. To address this gap, we propose a temporal modulation framework that fine-tunes LLMs and VLMs to effectively process time-series inputs alongside images and text, enabling VLA models to capture complex temporal dynamics and make more informed decisions in real-world robotic applications [3].

References:

1. Knowledge Inclusive Machine Learning for Disease Gene Prioritisation. CJ Gamage, Y Xia, R Rupasinghe, S Senevirathne, D Senanayake, …, S. Halgamuge. bioRxiv 2026.04.29.721522, 2026.
2. Knowledge Inclusive Adaptive Physics-Informed Neural Network for Microbial Interaction Modelling. R Rupasinghe, R Vidanaarachchi, A Hevapathige, S Seneviratne, S. L. Tang, S. Halgamuge. arXiv:2606.07686, 2026.
3. Rethinking Time Series Forecasting with LLMs via Nearest Neighbor Contrastive Learning. J Bogahawatte, S Seneviratne, M Perera, S Halgamuge. arXiv:2412.04806, 2026.