LETTER FROM THE EDITOR

Dear Colleague,

In this edition, we’ll cover an edge AI application domain that affects all of us: healthcare. Specifically, we’ll see how computer vision and agentic AI are performing real-time monitoring to transform our physical and mental health, and those of our elders, for example in detecting cognitive decline. We will also explore two takes on the near future of edge AI. But first…

We’re excited to announce our 2026 Embedded Vision Summit keynote speakers: Eric Xing, President of the Mohamed bin Zayed University of Artificial Intelligence, and Vikas Chandra, Senior Director at Meta Reality Labs.

Professor Xing will present recent breakthroughs in world models, fully open foundation models and parameter-efficient reasoning models. In addition to his position at the Mohamed bin Zayed University of Artificial Intelligence, he is a Professor of Computer Science at Carnegie Mellon University. His main research interests are in the development of machine learning and statistical methodology, as well as large-scale distributed computational systems and architectures, for solving problems involving automated learning, reasoning and decision-making in artificial, biological and social systems. In recent years, he has been focused on building large language models, world models, agent models and foundation models for biology. 

Vikas Chanda’s keynote, “Scaling Down Is the New Scaling Up,” will argue that the next decade will be about scaling down: AI that runs on your device, reasons across what you see and hear, and understands by utilizing context that never leaves your pocket. At Meta, Dr. Chandra leads an AI research team building efficient on-device AI for glasses and other mixed-reality products. These devices perceive the world as the wearer does, using context to anticipate needs and take action, laying the foundation for the next generation of human-device interaction. Prior to joining Meta in 2018, Dr. Chandra was Director of Applied Machine Learning at Arm Research, where his team helped pioneer techniques that enable AI to run on small, resource-constrained devices.
Check out our sessions and speakers, peruse the available event pass options, and then register today for the Summit, taking place May 11-13 in Santa Clara, California, using discount code 26EVSUM-NL for 25% off. We look forward to seeing you there!

Without further ado, let’s get to the content.

Erik Peters
Director of Ecosystem and Community Engagement, Edge AI and Vision Alliance

AI AND VISION ADVANCES IN HEALTHCARE

Virtual Reality, Machine Learning and Biosensing Advances Converging to Transform Healthcare and Beyond

In this wide-ranging interview, Walter Greenleaf, Neuroscientist at Stanford University’s Virtual Human Interaction Lab, explains how advances in virtual and augmented reality, machine learning and agentic AI and biosensing and embedded vision are converging to transform not only healthcare but human interaction as well. He details how this convergence will impact clinical care, disability solutions and personal health and wellness. Through real-time monitoring of physiological measurements, eye movements, voice tone, facial expressions and behavioral patterns, these integrated technologies are enabling sophisticated systems capable of sensing, analyzing and adapting to our arousal levels, cognitive status and emotional state, adjusting to individual preferences and interaction styles. Greenleaf examines how this technological revolution will transform physical and mental health as well as how humans interact with each other and with the world around us. You’ll learn how agentic AI and immersive visualization will unleash truly personalized experiences that reflect and enhance an individual’s physical and mental health.

Using Computer Vision for Early Detection of Cognitive Decline via Sleep-wake Data

AITCare-Vision predicts cognitive decline by analyzing sleep-wake disorder data in older adults. Using computer vision and motion sensors coupled with AI algorithms, AITCare-Vision continuously monitors sleep patterns, including disturbances such as frequent nighttime awakenings or irregular sleep cycles. AITCare-Vision utilizes this data to identify patterns that may signal cognitive decline, such as changes in sleep consistency or increased time spent awake at night. These insights are compared with baseline data to detect subtle shifts in cognitive health over time. In this presentation, Ravi Kota, CEO of AI Tensors, discusses the development of AITCare-Vision. He focuses on some of the key challenges his company addressed in the development process, including devising techniques to obtain accurate sleep-wake data without the use of wearables, designing the system to preserve privacy and implementing techniques to enable running AI models at the edge with low power consumption.

WHAT’S NEXT IN EDGE AI

On-Device LLMs in 2026: What Changed, What Matters, What’s Next

In this article, Vikas Chandra (a 2026 Embedded Vision Summit keynote speaker) and Raghuraman Krishnamoorthi explain why on-device LLMs on phones have shifted from “toy demos” to practical engineering—driven less by faster chips than by new approaches to model design, training, compression and deployment. They frame the motivation as four concrete benefits—lower latency, stronger privacy, lower serving cost and offline availability—while noting that frontier reasoning and very long conversations still tend to favor the cloud. They argue the binding constraint on phones is memory bandwidth (not TOPS), so 4-bit quantization and careful memory management (including KV-cache techniques) disproportionately improve real token throughput and usability under tight RAM and power limits. The authors then survey the “practical toolkit” (quantization, KV-cache strategies, speculative decoding, pruning) and increasingly mature deployment stacks (e.g., ExecuTorch, llama.cpp, MLX), and close by flagging what’s next: mixture-of-experts remains memory-movement-limited on edge, while test-time compute and on-device personalization look like major levers.

Edge AI and Vision at Scale: What’s Real, What’s Next, What’s Missing?

Edge AI and vision are no longer science projects—some applications, such as automotive safety systems, have already achieved massive scale. But for every success story, there are many more edge AI and computer vision products that have struggled to move beyond pilot deployments. So what’s holding them back? Scaling edge AI involves far more than just getting a model to run on a device. Challenges range from physical installation and fleet management to model updates, data drift, hardware changes and supply chain disruptions. And as systems grow, so do the variations in environments, sensor quality and real-world conditions. What does “scale” really mean in this space—and what does it take to get there? Exploring these questions is a panel of experts with firsthand experience deploying edge AI at scale, for a candid and practical discussion of what’s real, what’s next and what’s still missing.

Sally Ward-Foxton, Senior Reporter at EE Times, moderates our panel, featuring: Chen Wu, Director and Head of Perception at Waymo, Vikas Bhardwaj, Director of AI in the Reality Labs at Meta, Vaibhav Ghadiok, Chief Technology Officer of Hayden AI, and Gérard Medioni, Vice President and Distinguished Scientist at Amazon Prime Video and MGM Studios.

UPCOMING INDUSTRY EVENTS

Enabling Reliable Industrial 3D Vision with iToF Technology

 – e-con Systems Webinar: February 19, 2026, 11:00 am CET

MIPI CSI-2 over D-PHY & C-PHY: Advancing Imaging Conduit Solutions
– MIPI Alliance Webinar: February 24, 2026, 9:00 am PT

Robotics Builders Forum

 – February 25, Pittsburgh, Pennsylvania, 8:15 am – 5:30 pm ET. 

Cleaning the Oceans with Edge AI: The Ocean Cleanup’s Smart Camera Transformation

 – The Ocean Cleanup Webinar: March 3, 2026, 9:00 am PT

Why your Next AI Accelerator Should Be an FPGA

 – Efinix Webinar: March 17, 2026, 9:00 am PT

Embedded Vision Summit: May 11-13, 2026, Santa Clara, California

Newsletter subscribers may use the code 26EVSUM-NL for 25% off the price of registration.

FEATURED NEWS

Texas Instruments TDA5 Virtualizer Development Kit is accelerating next-generation automotive designs

Qualcomm, D3 Embedded and others will host Robotics Builders Forum, offering hardware, know-how and networking

Microchip has extended its edge AI offering with full-stack solutions that streamline development 

More News

Dear Colleague,

Whether you’re at one of the big AI players making headlines, or trying to break out with a startup, many of our readers are on their own journey to scale—turning prototypes into robust products, moving from research workflows into production pipelines, and scaling deployments in the real world. We’ll hear perspectives on scaling from both business leaders and technical experts. But first, I’d like to share a few exciting updates from the Alliance.

On Tuesday, March 17, the Edge AI and Vision Alliance is pleased to present a webinar in collaboration with Efinix. Edge AI system developers often assume that AI workloads require a GPU or NPU. But when cost, latency, complex I/O or tight power budgets dominate, FPGAs offer compelling advantages. Mark Oliver, VP of Marketing and Business Development at Efinix, explores how FPGAs serve not just as a compute block, but as a system-integration and acceleration platform that can combine tailored sensor I/O, signal processing, pre/post-processing and neural inference on one device. Mark will also show how to map AI models onto FPGAs without doing custom hardware design, using two two practical on-ramps—(1) a software-first flow that generates custom instructions callable from C, and (2) a turnkey CNN acceleration block. More info here.

We’re also excited to announce our first batch of expert speakers and sessions for the 2026 Embedded Vision Summit. These speakers will soon be joined by dozens more, all focused on building products using computer vision and physical AI, so stay tuned! The Embedded Vision Summit returns to Santa Clara, California May 11-13.

Without further ado, let’s get to the content.

Erik Peters
Director of Ecosystem and Community Engagement, Edge AI and Vision Alliance

FROM PROTOTYPE TO OPERATIONS

Deep Sentinel: Lessons Learned Building, Operating and Scaling an Edge AI Computer Vision Company 

Deep Sentinel’s edge AI security cameras stop some 45,000 crimes per year. Unlike most security camera systems, they don’t just record video for later playback: they use edge AI, vision and humans in the loop to detect crimes in progress. And then they react—quickly!—to stop the bad guys. In this humorous and fast-paced talk, David Selinger, CEO of Deep Sentinel, shares some hard lessons he learned in his journey taking Deep Sentinel’s AI cameras from idea to product. From the perspective of a software guy trying to build hardware, you’ll hear about pitfalls ranging from the challenges of low-volume manufacturing to the joys of hardware vendor software support. If you’re bringing a vision product to market, you can’t afford to miss this presentation—and if you’re a hardware, software or services supplier, come learn what you can do to make your customers’ lives easier.

Taking Computer Vision Products from Prototype to Robust Product 

When developing computer vision-based products, getting from a proof of concept to a robust product ready for deployment can be a massive undertaking. The most vexing challenges in this process often relate to the “long-tail problem,” which arises when datasets have highly imbalanced distributions of classes. This candid conversation between Chris Padwick, Machine Learning Engineer at Blue River Technology, and Mark Jamtgaard, Director of Technology at RetailNext, focuses on the realities of delivering reliable computer vision products to market, delves into lessons learned from Padwick’s years of experience developing automated farming equipment for deployment at scale and explores practical strategies for data curation, data labeling and model testing approaches. Padwick and Jamtgaard also discuss approaches for tackling challenges such as object class confusion and correlated training data.

SCALING THE TECHNICAL STACK

Scaling Computer Vision at the Edge

In this presentation, Eric Danziger, CEO of Invisible AI, introduces a comprehensive framework for scaling computer vision systems across three critical dimensions: capability evolution, infrastructure decisions and deployment scaling. Today’s leading-edge vision systems leverage scalable models that, when utilized through prompting, enable advanced capabilities without the resource demands of general-purpose AI vision. However, scaling these systems faces significant edge computing challenges, where limited compute power and networking capabilities restrict the number of camera streams that can be processed, leading to increased costs and complexity. Danziger presents a structured approach to navigating these trade-offs, showcasing automation tools and deployment strategies that help engineering teams with limited resources maximize capabilities while making optimal decisions between edge and cloud processing architectures.

Scaling Machine Learning with Containers: Lessons Learned

In the dynamic world of machine learning, efficiently scaling solutions from research to production is crucial. In this presentation, Rustem Feyzkhanov, Machine Learning Engineer at Instrumental, explores the nuances of scaling machine learning pipelines, emphasizing the role of containerization in improving reproducibility, portability and scalability. Key topics include building efficient training pipelines, monitoring models in production and optimizing costs while handling peak loads. You’ll learn practical strategies for bridging the gap between research and production, ensuring consistent performance and rapid iteration cycles. Tailored for professionals, this presentation delivers actionable insights to enhance the scalability and robustness of ML systems across diverse applications.

UPCOMING INDUSTRY EVENTS

Cleaning the Oceans with Edge AI: The Ocean Cleanup’s Smart Camera Transformation

 – The Ocean Cleanup Webinar: March 3, 2026, 9:00 am PT

Why your Next AI Accelerator Should Be an FPGA

 – Efinix Webinar: March 17, 2026, 9:00 am PT

Embedded Vision Summit: May 11-13, 2026, Santa Clara, California
Newsletter subscribers may use the code 26EVSUM-NL for 25% off the price of registration.

FEATURED NEWS

NAMUGA has launched the Stella-2 next-generation 3D LiDAR sensor

Google has added “Agentic Vision” to Gemini 3 Flash

Yole Group discusses why DRAM prices keep rising in the age of AI

Microchip has expanded the PolarFire FPGA Smart Embedded Video ecosystem with new SDI IP cores and a quad CoaXPress bridge kit

NanoXplore and STMicroelectronics have delivered a european FPGA for space missions

More News

LETTER FROM THE EDITOR

Dear Colleague,

On Tuesday, March 3, the Edge AI and Vision Alliance is pleased to present a webinar in collaboration with The Ocean Cleanup. The Ocean Cleanup is on a mission to rid the world’s oceans of plastic. To do that, the team needs to know where plastic accumulates, how it moves, and how their cleanup systems behave in tough, remote marine environments. Robin de Vries, Lead for Autonomous Debris Imaging System (ADIS) will walk attendees through their development, from the first generation of GoPros and removable hard drives to their current setup: a customized smart camera platform that runs computer vision models on the device. Robin will discuss system design for marine environments, hardware choices, power and thermal limits, model deployment and remote management, as well as tradeoffs and lessons learned. More info here.

This issue, we’ll conclude our two-part feature on foundational vision/AI techniques, and we’ll touch on one of the applications that always receives a lot of attention at CES: autonomous driving. Frank Moesle from Valeo provides both business insights on software-defined vehicles (SDVs), sensor fusion, and software reliability, as well as technical insights into ADAS for SDVs. If you enjoy Frank’s perspectives, he’s confirmed to return to this year’s Embedded Vision Summit, May 11-13 in Santa Clara, California.

Without further ado, let’s get to the content.

Erik Peters
Director of Ecosystem and Community Engagement, Edge AI and Vision Alliance

COMPUTER VISION MODEL FUNDAMENTALS

Transformer Networks: How They Work and Why They Matter

Transformer neural networks have revolutionized artificial intelligence by introducing an architecture built around self-attention mechanisms. This has enabled unprecedented advances in understanding sequential data, such as human languages, while also dramatically improving accuracy on nonsequential tasks like object detection. In this talk, Rakshit Agrawal, formerly Principal AI Scientist at Synthpop AI, explains the technical underpinnings of transformer architectures, from input data tokenization and positional encoding to the self-attention mechanism, which is the core component of these networks. He also explores how transformers have influenced the direction of AI research and industry innovation. Finally, he touches on trends that will likely influence how transformers evolve in the near future.

Understanding Human Activity from Visual Data

Activity detection and recognition are crucial tasks in various industries, including surveillance and sports analytics. In this talk, Mehrsan Javan, Chief Technology Officer at Sportlogiq,  provides an in-depth exploration of human activity understanding, covering the fundamentals of activity detection and recognition, and the challenges of individual and group activity analysis. He uses examples from the sports domain, which provides a unique test bed requiring analysis of activities involving multiple people, including complex interactions among them. Javan traces the evolution of technologies from early deep learning models to large-scale architectures, with a focus on recent technologies such as graph neural networks, transformer-based models, spatial and temporal attention and vision-language approaches, including their strengths and shortcomings. Additionally, he examines the computational and deployment challenges associated with dataset scale, annotation complexity, generalization and real-time implementation constraints. He concludes by outlining potential challenges and future research directions in activity detection and recognition.

AUTONOMOUS DRIVING & ADAS

Three Big Topics in Autonomous Driving and ADAS

In this on-stage interview, Frank Moesle, Software Department Manager at Valeo, and independent journalist Junko Yoshida focus on trends and challenges in automotive technology, autonomous driving and ADAS. First up: Sensor fusion is often touted as the perception solution for autonomy. But what exactly is it? What’s involved and what are the challenges? Next, Moesle and Yoshida discuss the trend toward “software-defined everything” in automotive. Is it just a buzzword, or are there places where it brings real value? And finally, they touch on software reliability: if cars are becoming increasingly autonomous and dependent on software, how do we build automotive systems that are safe and reliable?

Toward Hardware-agnostic ADAS Implementations for Software-defined Vehicles

ADAS (advanced-driver assistance systems) software has historically been tightly bound to the underlying system-on-chip (SoC). This software, especially for visual perception, has been extensively optimized for specific SoCs and their dedicated accelerators. In this talk, Frank Moesle, Software Department Manager at Valeo, explains the historic reasons for this approach and shows its advantages. Recent developments, however, such as the emergence of middleware solutions, allow the decoupling of embedded software from the hardware and its specific accelerators, enabling the creation of true software-defined vehicles. Moesle explains how such an approach can achieve efficient implementations, including the use of emulation and cloud processing, and how this benefits not only Tier 1 automotive subsystem suppliers, but also SoC vendors and auto manufacturers.

UPCOMING INDUSTRY EVENTS

Cleaning the Oceans with Edge AI: The Ocean Cleanup’s Smart Camera Transformation

 – The Ocean Cleanup Webinar: March 3, 2026, 9:00 am PT

Embedded Vision Summit: May 11-13, 2026, Santa Clara, California

Newsletter subscribers may use the code 26EVSUM-NL for 25% off the price of registration.

FEATURED NEWS

Qualcomm’s has expanded its IoT edge AI offerings developers, enterprises & OEMs

Ambarella has launched a powerful 8K Vision AI SoC with and multi-sensor perception performance

NVIDIA has released the Jetson T4000 and NVIDIA JetPack 7.1 for edge inference

NXP has introduced its eIQ agentic AI framework for autonomous intelligence at the edge

ModelCat AI is delivering rapid ML model onboarding in partnership with Alif Semiconductor

Chips&Media and Visionary.ai have unveiled the world’s first AI-based full image signal processor

LETTER FROM THE EDITOR

Dear Colleague,

Happy New Year and Happy CES for those attending! If you’re heading to CES and want to see the latest in physical AI and computer vision technologies, check out our Directory of Alliance Members at CES to see what they are showing, where to find them and easy ways to set up appointments for suite and demo visits. 

Would you or a colleague benefit from an introduction to (or review) of foundational vision/AI techniques? If so, check out the insightful presentations below from experienced practitioners.  (And stay tuned for more in our next Insights newsletter edition in two weeks.)

We’ll also catch up on some news highlights you may have missed over the holidays.

Lastly, don’t miss your chance to save 25% on registration for the 2026 Embedded Vision Summit, coming up May 11-13 in Santa Clara, California! Register with code 26EVSUM-NL to take advantage of this price.

Without further ado, let’s dig in!

Erik Peters
Director of Ecosystem and Community Engagement, Edge AI and Vision Alliance

DEEP LEARNING LITERACY: BACK TO BASICS

Introduction to Deep Learning and Visual AI: Fundamentals and Architectures

This talk provides a high-level introduction to artificial intelligence and deep learning, covering the basics of machine learning and the key concepts of deep learning. Mohammad Haghighat, Senior Manager for CoreAI at eBay, explores the different types of deep learning architectures, including fully connected networks, convolutional neural networks (CNNs), recurrent neural networks (RNNs), 3D CNNs and transformers, highlighting their most common use cases and applications. He then focuses on visual AI, introducing CNNs as a fundamental architecture for image and video analysis. Haghighat discusses the building blocks of CNNs and explores example architectures such as Inception, ResNet and EfficientNet. Finally, he highlights some recent trends in visual AI such as vision transformers (ViTs), hybrid architectures and vision-language models (VLMs). You will gain a solid understanding of the fundamentals of deep learning and visual AI, as well as recent advancements and current trends in the field.

Introduction to DNN Training: Fundamentals, Process and Best Practices

Training a model is a crucial step in machine learning, but it can be overwhelming for beginners. In this talk, Kevin Weekly, CEO of Think Circuits, provides a comprehensive introduction to the fundamentals of model training. He introduces the different types of training, such as supervised, unsupervised and semi-supervised learning, and then delves into techniques for supervised training. He explains the training process, including error surfaces, optimization methods and back-propagation. Weekly explains key concepts such as trainable parameters and data requirements. He also discusses the main “knobs” that control the training process, such as hyperparameters, regularization and batch normalization, and provides an overview of metrics to monitor during training, including loss curves, model accuracy and precision. Additionally, he covers common problems that arise during training, such as overfitting and underfitting, and introduces approaches to address these issues. Finally, he touches on popular training frameworks and provides resources for further learning.

DATA, PIPELINES & MLOPS

Mastering the End-to-end Machine Learning Model Building Process: Best Practices and Pitfalls

In this talk, Paril Ghori, Senior Data Scientist at Caterpillar, explores the complete machine learning model building process, providing data scientists and ML engineers with practical insights and strategies for success. He examines each phase of the model life cycle, from data ingestion and pre-processing to feature engineering, model selection, training and fine-tuning. He explains best practices for model evaluation, validation and deployment, including effective MLOps integration to ensure seamless model monitoring and scalability. Ghori highlights real-world case studies and common pitfalls encountered during model development, offering actionable solutions and strategies to overcome challenges. He emphasizes optimizing workflows to improve performance and ensure reproducibility in complex projects. You’ll gain a deeper understanding of the entire model building process and gain insights that will help you build robust, efficient and scalable models to drive impactful business outcomes and support continuous innovation.

Multimodal Enterprise-scale Applications in the Generative AI Era

As artificial intelligence is making rapid strides in use of large language models, the need for multimodality arises in multiple application scenarios. Similar to the way humans use multiple sensory systems to solve problems and arrive at decisions, in many applications AI problem-solving is enriched by using multimodal inputs. In this presentation, Mumtaz Vauhkonen, Senior Director of AI at Skyworks Solutions, explores the process of building multimodal applications at scale, focusing on the core aspects of quality dataset creation, multimodal data fusion techniques and model pipelines for enterprise applications. She also examines the challenges that arise in bringing these applications to production and techniques for addressing these challenges.

UPCOMING INDUSTRY EVENTS

Embedded Vision Summit: May 11-13, 2026, Santa Clara, California

FEATURED NEWS

NVIDIA has debuted the Nemotron 3 family of open models, featuring a hybrid latent MoE architecture

poLight ASA and Image Quality Labs have announced an M12-based RPi TLens development platform for rapid evaluation of high speed, constant field-of-view focusing functionality

NVIDIA’s agreement with Groq will accelerate AI inference at global scale

Samsung is rumored to be launching an in-house mobile GPU by 2027

Google has released FunctionGemma, a lightweight function-calling model aimed at on-device agents

More News

LETTER FROM THE EDITOR

Dear Colleague,
Welcome to our annual CES Special Edition. We’ll give you a rundown on some of the most interesting companies to see  at CES and we’ll highlight some consumer-facing applications of edge AI. Then we’ll continue our focus on technical content with two presentations addressing challenges in developing consumer edge AI products, and we’ll round it out with a recap of the news. Let’s dig in!

The Edge AI and Vision Alliance is proud to showcase companies that will be presenting the latest computer vision and AI technologies at CES. Check out our Directory of Alliance Members at CES to see what they are showing, where to find them and easy ways to set up appointments for suite and demo visits. 

Edge AI and Vision Technology Companies to See at CES 2026:

• AIRY3D
• AMD
• Arm
• Au-Zone
• Axelera AI
• Cadence
• CLIKA
• Commonlands
• Digica
• e-con Systems
• ENERZAi
• Expedera
• eYs3D Microelectronics
• FotoNation
• Inuitive
• Magik Eye
• MemryX
• Microchip Technology
• NAMUGA
• Nextchip
• Nota AI
• NXP Semiconductors
• Quadric
• Rapidflare
• STMicroelectronics
• Symage by Geisel Software
• TekStart
• Texas Instruments
• VeriSilicon
• videantis
• Visidon
• Visionary.ai

I’m also pleased to announce that registration for the 2026 Embedded Vision Summit is at its lowest price, 35% off, from now through December 31! The Summit will take place May 11-13 in Santa Clara, California, and we very much hope to see all of you there. 

If you’d like to present at the 2026 Summit, our Call for Presentation Proposals also remains open, as we’ve extended the deadline to Friday, December 19. Check out the 2026 topics list on the Call for Proposals page, and submit your proposal today.

Erik Peters
Director of Ecosystem and Community Engagement, Edge AI and Vision Alliance

VISION DEPLOYED CONSUMER PRODUCTS

Enabling Ego Vision Applications on Smart Eyewear Devices 

Ego vision technology is revolutionizing the capabilities of smart eyewear, enabling applications that understand user actions, estimate human pose and provide spatial awareness through simultaneous localization and mapping (SLAM). This presentation dives into the latest advancements in deploying these computer vision techniques on embedded systems. Francesca Palermo, Research Principal Investigator at EssilorLuxottica, explains how her company overcomes the challenges of constrained processing power, memory and energy consumption while still achieving real-time, on-device performance for smart eyewear. In particular, she shares insights on optimizing neural networks for low-power environments, innovating in pose estimation and effectively integrating SLAM in dynamic settings, all supported by real-world examples and demonstrations. She also explores how these capabilities open new possibilities for augmented reality, assistive technologies and enhanced personal health.

AI-powered Scouting: Democratizing Talent Discovery in Sports

In this presentation, Jonathan Lee, Chief Product Officer at ai.io, shares his company’s experience using AI and computer vision to revolutionize talent identification in sports. By developing aiScout, a platform that enables athletes to upload drill videos for AI evaluation, ai.io aims to democratize access to scouting. Leveraging 3DAT, their AI-driven biomechanics analysis tool, they extract precise movement data without sensors or wearables. Lee discusses how AI-driven scouting levels the playing field, allowing athletes to get discovered based on ability, not access—proven with elite athletes from the Tokyo Olympics to the NFL Scouting Combine. He also covers the business model, scalability and future of AI-driven scouting, highlighting its potential to redefine talent discovery and development.

KNOWLEDGE DISTILLATION AND OBJECT DETECTORS

Introduction to Knowledge Distillation: Smaller, Smarter AI Models for the Edge

As edge computing demands smaller, more efficient models, knowledge distillation emerges as a key approach to model compression. In this presentation, David Selinger, CEO of Deep Sentinel, delves into the details of this process, exploring what knowledge distillation entails and the requirements for its implementation, including dataset size and tools. Selinger examines when to use knowledge distillation, its pros and cons, and showcases examples of successfully distilled models. Based on performance data highlighting the benefits of distillation, he concludes that knowledge distillation is a powerful tool for creating smaller, smarter models that thrive at the edge.

Object Detection Models: Balancing Speed, Accuracy and Efficiency

Deep learning has transformed many aspects of computer vision, including object detection, enabling accurate and efficient identification of objects in images and videos. However, choosing the right deep neural network-based object detector for your project, particularly when deploying on lightweight hardware, requires consideration of trade-offs between accuracy, speed and computational efficiency. In this talk, Sage Elliott, AI Engineer at Union.ai, introduces the fundamental types of DNN-based object detectors. He covers models such as Faster R-CNN for high-accuracy applications and single-stage models such as YOLO and SSD for faster processing. He discusses lightweight architectures, including MobileNet, EfficientDet and vision transformers, which optimize object detection for resource-constrained environments. You will learn the trade-offs between object detection models for your computer vision applications, enabling informed choices for optimal performance and deployment.

UPCOMING INDUSTRY EVENTS

AI Everywhere 2025 – EE Times Virtual Event: December 10-11, 2025

 Embedded Vision Summit: May 11-13, 2026, Santa Clara, California

FEATURED NEWS

Qualcomm has released the premium tier Snapdragon 8 Gen 5, driving performance and new user experiences

AMD has released its Spartan UltraScale+ FPGA SCU35 Evaluation Kit, and announced Infineon HyperRAM support on the platform

Intel has broadened support for LLMs and VLMs with the release of OpenVINO 2025.4

NVIDIA and Synopsys have announced a strategic partnership to revolutionize engineering and design through a raft of initiatives

Chips&Media’s WAVE-N v2 Custom NPU delivers higher TOPS and greater power efficiency

More News

LETTER FROM THE EDITOR

Dear Colleague,

Good morning. Before we get to the content today, I’d like to introduce myself, since you’ll be hearing from me regularly. I’m Erik Peters, the new editor of the Edge AI and Vision Insights newsletter. I’ve been helping the Alliance with market and industry ecosystem research (and the occasional ML project) for over seven years, and I’m pleased to have a new opportunity to bring some of that work directly to you. Each newsletter, we’ll showcase some outstanding technical content, and we’ll see some real-world applications as well. Without further ado, let’s dive in!

On Thursday, December 11, Jeff Bier, Founder of the Edge AI and Vision Alliance will present at two sessions as part of EE Times’ AI Everywhere 2025. In the first, “Edge AI Everywhere: What Will It Take To Get There?” Jeff will highlight some inspiring success stories and explore the key challenges that need to be overcome to enable more edge AI-based systems to reach massive scale. In the second, Jeff will be joined by speakers from Alliance Member companies Ambarella, STMicroelectronics, and Synopsys for a panel discussion, “Deploying and Scaling AI at the Edge – From Lab to Life Cycle.” More information on the event and how to register can be found here.

I’m also pleased to announce that registration for the 2026 Embedded Vision Summit is now open! The Summit will take place May 11-13 in Santa Clara, California, and we very much hope to see all of you there.

Our Call for Presentation Proposals for the 2026 Summit also remains open through December 5. We’re planning more than 100 expert sessions and would love to see your ideas—from physical AI case studies to efficient edge AI techniques to the latest advances in vision language models. Check out the 2026 topics list on the Call for Proposals page for inspiration and to submit your own proposal by December 5.

Erik Peters
Director of Ecosystem and Community Engagement, Edge AI and Vision Alliance

BUILDING AND DEPLOYING REAL-WORLD ROBOTS

Real-world Deployment of Mobile Material Handling Robotics in the Supply Chain

More and more of the supply chain needs to be, and can be, automated. Demographics, particularly in the developed world, are driving labor scarcity. Additionally, in manual material handling turnover, injury rates and absenteeism are rampant. Fortunately, modern warehouse robotic systems are becoming able to see and manipulate cartons and bags at levels of speed and dependability that can deliver strongly positive ROI to supply chain operators. However, moving such systems into production in the complexity of real-world warehouses requires exceptional levels of product capability and rigor around testing and deployment practices. Many equipment vendors reach the pilot stage, then fail to break through to production due to the absence of this rigor. Peter Santos, Chief Operating Officer of Pickle Robot Company, focuses on three core principles of successful production deployments: complexity acceptance, test set design and early customer collaboration.

Integrating Cameras with the Robot Operating System (ROS)

In this presentation, Karthik Poduval, Principal Software Development Engineer at Amazon Lab126, explores the integration of cameras within the Robot Operating System (ROS) for robust embedded vision applications. He delves into ROS’s core functionalities for camera data handling, including the ROS messages (data structures) used for transmitting image data and calibration parameters. Poduval discusses essential camera calibration techniques, highlighting the importance of determining accurate intrinsic and extrinsic parameters. He also explains open-source ROS nodes, such as those within image_proc and stereo_image_proc, that facilitate crucial post-processing steps, including distortion correction and rectification. The presentation equips you with practical knowledge to leverage ROS’s capabilities for building advanced vision-enabled robotic systems.

DESIGNING APPLICATION-SPECIFIC CAMERA SYSTEMS

Specifying and Designing Cameras for Computer Vision Applications

Designing a camera system requires a deep understanding of the fundamental principles of image formation and the physical characteristics of its components. Translating computer vision-based system requirements into camera system parameters is a crucial step. In this talk, Richard Crisp, Vice President and CTO at Etron Technology America, provides a comprehensive overview of the key concepts involved in designing or specifying a camera system. Crisp covers the basics of image formation and the associated physics. He discusses sensor and lens effects that impact image quality, such as diffraction, circle of confusion, and depth of field. He also discusses application-specific requirements, including lighting conditions, frame rate and motion blur. Finally, he presents a detailed example illustrating how to translate application requirements into camera parameters, highlighting cost-performance trade-offs. You will gain a thorough understanding of the key factors influencing camera system design and be able to make informed decisions when selecting or designing a camera system.

Developing a GStreamer-based Custom Camera System for Long-range Biometric Data Collection

In this presentation, Gavin Jager, Researcher and Lab Space Manager at Oak Ridge National Laboratory, describes Oak Ridge National Laboratory’s work developing software for a custom camera system based on GStreamer. The BRIAR project requires high-quality video capture at distances over 400 meters for biometric recognition and identification, but commercial cameras struggle to capture high-quality video at these distances. To address this, Oak Ridge National Laboratory developed a custom camera system using GStreamer, enabling advanced imaging capabilities and long-range data capture. The work included designing a GStreamer pipeline capable of managing multiple sensor formats, integrating UDP server hooks to manage recording, using GstRTSPServer to build an RTSP server and creating an extensible hardware control interface. By integrating with network video recorders, Oak Ridge National Laboratory simplified monitoring, data handling and curation, and successfully supported BRIAR’s complex data collection efforts. This presentation details the GStreamer-based implementation, highlighting technical challenges faced and how they were overcome.

UPCOMING INDUSTRY EVENTS

AI Everywhere 2025 – EE Times Virtual Event: December 10-11, 2025

Embedded Vision Summit: May 11-13, 2026, Santa Clara, California

FEATURED NEWS

Micron has shipped its automotive universal flash storage (UFS) 4.1 for enhanced ADAS and cabin experience

Au-Zone Technologies has expanded access to EdgeFirst Studio, an MLOps platform for spatial perception at the edge

Vision Components has enabled support for its MIPI Cameras on the SMARC IMX8M Plus Development Kit from ADLINK

Axelera has released its Metis PCIe with 4 Quad-Core AIPUs

More News

EDGE AI AND VISION PRODUCT OF THE YEAR WINNER SHOWCASE

LETTER FROM THE EDITOR

Dear Colleague,

Next Tuesday, November 18 at 9 am PT, the Yole Group will present the free webinar “How AI-enabled Microcontrollers Are Expanding Edge AI Opportunities” in partnership with the Edge AI and Vision Alliance. Running AI inference at the edge, versus in the cloud, has many compelling benefits; greater privacy, lower latency and real-time responsiveness key among them. But implementing edge AI in highly cost-, power-, or size-constrained devices has historically been impractical due to its compute, memory and storage implementation requirements.

Nowadays, however, the AI accelerators and related resources included in modern microcontrollers, in combination with technology developments and toolset enhancements that shrink the size of deep learning models, are making it possible to run computer vision, speech interfaces, and other AI capabilities at the edge.

In this webinar, Tom Hackenberg, Principal Analyst for Computing at the Yole Group, will explain that while scaling AI upward into massive data centers may dominate today’s headlines, scaling downward to edge applications may be even more transformative. Hackenberg will share market size and forecast data, along with supplier product and developer application case study examples, to support his contention that edge deployment is key to unlocking AI’s full potential across industries. A question-and-answer session will follow the presentation. For more information and to register, please see the event page.

Erik Peters
Director of Ecosystem and Community Engagement, Edge AI and Vision Alliance

DESIGN CONSIDERATIONS FOR ROBOTICS APPLICATIONS

Lessons Learned Building and Deploying a Weed-killing Robot
Agriculture today faces chronic labor shortages and growing challenges around herbicide resistance, as well as consumer backlash to chemical inputs. Smarter, more sustainable approaches are needed to secure the ongoing production of fresh produce. In this session, Xiong Chang, CEO and Co-founder of Tensorfield Agriculture, introduces his company’s unique robot, which uses high-speed computer vision to enable extremely precise, pesticide-free robotic weeding. Chang highlights some of the key challenges his team faced in developing the robot. He explains Tensorfield’s business model, shows how its technology has the potential to save millions of dollars in labor and material costs, and shares how Tensorfield plans to scale its business.

Sensors and Compute Needs and Challenges for Humanoid Robots
Vlad Branzoi, Perception Sensors Team Lead at Agility Robotics, presents the “Sensors and Compute Needs and Challenges for Humanoid Robots” tutorial at the September 2025 Edge AI and Vision Innovation Forum.

AGENTIC AI AT THE EDGE

Introduction to Designing with AI Agents
Artificial intelligence agents are components in an AI system that can perform tasks autonomously, making decisions and taking actions on their own. In this talk, Frantz Lohier, Senior Worldwide Specialist for Advanced Computing, AI and Robotics at Amazon Web Services, explores the concept of AI agents, their benefits and how they can revolutionize AI development. He discusses the differences between agentic and non-agentic workflows, and how agents can improve the performance of existing models through reflection, tool use, planning and multiagent collaboration. Lohier examines the types of AI agents, such as vision agents, LLM agents, math solver agents and code generation agents, and how they can be used in various AI-based systems. He also discusses how agents are created, trained, tested and integrated into AI systems. You’ll gain a working understanding of AI agents, their benefits and how they can be used in AI-based application development.

Building Agentic Applications for the Edge
Along with AI agents, the new generation of large language models, vision-language models and other large multimodal models are enabling powerful new capabilities that promise to transform industries. In this talk, Amit Mate, Founder and CEO of GMAC Intelligence, explores the requirements and architectures of agentic applications, including AI and non-AI requirements, and explores two main approaches to agent-based application architecture: integrating separate models and multimodal approaches. Through detailed examples, he demonstrates the pros and cons of each approach and discusses the challenges and opportunities of building practical agent-based applications on edge devices, including challenges associated with implementing large models at the edge.

UPCOMING INDUSTRY EVENTS

How AI-enabled Microcontrollers Are Expanding Edge AI Opportunities – Yole Group Webinar: November 18, 2025, 9:00 am PT

Embedded Vision Summit: May 11-13, 2026, Santa Clara, California

More Events

FEATURED NEWS

Axelera AI Announces the Europa AIPU, Setting New Industry Benchmark for AI Accelerator Performance, Power Efficiency and Affordability

STMicroelectronics Empowers Data-Hungry Industrial Transformation with a Unique Dual-Range Motion Sensor

NXP Semiconductors Completes the Acquisitions of Aviva Links and Kinara to Advance Automotive Connectivity and AI at the Intelligent Edge

Qualcomm Launches the AI200 and AI250: Redefining Rack-scale Data Center Inference Performance for the AI Era

BrainChip Unveils the Breakthrough AKD1500 Edge AI Co-Processor at Embedded World North America

More News

EDGE AI AND VISION PRODUCT OF THE YEAR WINNER SHOWCASE

Dear Colleague,

The Call for Presentation Proposals for the 2026 Embedded Vision Summit, taking place May 11-13 in Santa Clara, California, is now open! We’re planning more than 100 expert sessions and would love to see your ideas—from physical AI case studies to efficient edge AI techniques to the latest advances in vision language models. Check out the 2026 topics list on the Call for Proposals page for inspiration and to submit your own proposal, due December 5th!

Brian Dipert
Editor-In-Chief, Edge AI and Vision Alliance

DEPTH SENSING AND SLAM OPTIONS

Introduction to Depth Sensing: Technologies, Trade-offs and Applications

Depth sensing is a crucial technology for many applications, including robotics, automotive safety and biometrics. In this 2025 Embedded Vision Summit Presentation, Chris Sarantos, Independent Consultant with Think Circuits, provides an overview of depth sensing technologies, including stereo vision, time-of-flight (ToF) and LiDAR. He discusses the capabilities, trade-offs and limitations of each technology and explores how these fit with the requirements of typical applications. Sarantos examines key characteristics such as range, depth resolution, transverse resolution, frame rate and eye safety. He also discusses the latest advances in ToF sensors and LiDAR, including the use of single-photon avalanche diode (SPAD) arrays, optical phased arrays and chirped pulse techniques. He presents a comparison matrix summarizing the key characteristics and typical applications of each technology. This comprehensive introduction to depth sensing, including the strengths and weaknesses of each technology, will enable you to make informed decisions for your applications.

Optimizing Real-time SLAM Performance for Autonomous Robots with GPU Acceleration

Optimizing execution time of long-term and large-scale SLAM algorithms is essential for real-time deployments on edge compute platforms. Faster SLAM output means faster map refresh rates and quicker decision-making. RTAB-Map is a popular state-of-the-art SLAM algorithm used in autonomous mobile robots. RTAB-Map is implemented in an open-source library that supports various sensors, including RGB-D cameras, stereo cameras and LiDAR. In this 2025 Embedded Vision Summit talk, Naitik Nakrani, Solution Architect Manager at eInfochips, explains how LiDAR-based SLAM implemented with RTAB-Map can be accelerated by leveraging GPU-based libraries on NVIDIA platforms. He shares a detailed optimization methodology and results. He also shares effective ways in which SLAM algorithms can be accelerated on resource-constrained devices.

DEVELOPING AND IMPLEMENTING VISION-LANGUAGE MODELS

Vision-language Models on the Edge

In this 2025 Embedded Vision Summit presentation, Cyril Zakka, Health Lead at Hugging Face, provides an overview of vision-language models (VLMs) and their deployment on edge devices using Hugging Face’s recently released SmolVLM as an example. He examines the training process of VLMs, including data preparation, alignment techniques and optimization methods necessary for embedding visual understanding capabilities within resource-constrained environments. Zakka explains practical evaluation approaches, emphasizing how to benchmark these models beyond accuracy metrics to ensure real-world viability. And to illustrate how these concepts play out in practice, he shares data from recent work implementing SmolVLM in an edge device.

LLMs and VLMs for Regulatory Compliance, Quality Control and Safety Applications

By using vision-language models (VLMs) or combining large language models (LLMs) with conventional computer vision models, we can create vision systems that are able to interpret policies and enable a much more sophisticated understanding of scenes and human behavior compared with current-generation vision models. In this 2025 Embedded Vision Summit talk, Lazar Trifunovic, Solutions Architect at Camio, illustrates these capabilities with several examples of commercial applications targeting use cases such as ensuring compliance with safety policies and manufacturing regulations. He also shares the lessons his company has learned about the limitations and challenges of utilizing LLMs and VLMs in real-world applications.

UPCOMING INDUSTRY EVENTS

How AI-enabled Microcontrollers Are Expanding Edge AI Opportunities – Yole Group Webinar: November 18, 2025, 9:00 am PT

Embedded Vision Summit: May 11-13, 2026, Santa Clara, California

More Events

FEATURED NEWS

Synaptics Launches the Next Generation of Astra Multimodal GenAI Processors to Power the Future of the Intelligent IoT Edge

FRAMOS Unveils Three Specialized Camera Modules for UAV and Drone Applications

Qualcomm to Acquire Arduino, Accelerating Developers’ Access to its Leading Edge Computing and AI 

STMicroelectronics Introduces New Image Sensors for Industrial Automation, Security and Retail Applications

NVIDIA DGX Spark Arrives for World’s AI Developers

More News

EDGE AI AND VISION PRODUCT OF THE YEAR WINNER SHOWCASE