Edge AI and Vision Insights: August 19, 2026

LETTER FROM THE EDITOR

Dear Colleague,

This week’s issue looks at what it takes to make increasingly capable AI systems practical at the edge. We explore SLAM from foundational concepts through radar-based implementations for automotive and robotics and examine techniques for fitting sophisticated models into constrained devices through smaller language models, quantization and self-compression. In the news, we cover advances in open foundation models, edge AI tools and processors. And, as these technologies continue to evolve, we’d like to hear from the engineers putting them into practice.

If you’d like to present at the 2027 Embedded Vision Summit, good news: our Call for Presentation Proposals remains open, as we’ve extended the deadline to Friday, August 28. Check out the topics list on the Call for Proposals page and submit your proposal today.

Another great way to share your expertise is to participate in our annual Computer Vision and Physical AI Developer Survey. Many providers of the building-block technologies that enable computer vision and physical AI products and systems use the results of this annual survey to guide their priorities. We share the survey results at Alliance events, in white papers and presentations throughout the year on the Alliance website, and, of course, with everyone who completes the survey! Our target demographic for this survey is people who are or have recently been directly involved in an engineering role developing systems or applications using computer vision or other types of physical (sensor-based) AI. Take the Survey.

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

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

SLAM: FROM FUNDAMENTALS TO RADAR

Introduction to Simultaneous Localization and Mapping: From Block Diagrams to Real Systems

Simultaneous localization and mapping (SLAM) is the core capability that allows robots and autonomous systems to build a map of an unknown environment while estimating their own motion within it. In this talk, Amit Gupta, Associate Director–Solution Architecture and Head of Robotics CoE at eInfochips (an Arrow company), explains what SLAM is, why it matters, how it works and where it shows up in real products—from mobile robots and drones to AR/VR and automotive. He walks through a conceptual formulation and block diagram of a modern SLAM system, defines key terminology and describes the main components (front-end matching, state estimation, map management and loop closure) with minimal math. The session highlights practical design trade-offs such as compute versus accuracy, common algorithmic failure modes (dynamic scenes, sensor motion, degraded features) and implementation challenges in real-time systems. He concludes with sensor options and multimodal fusion, a visual example of SLAM in action and a survey of off-the-shelf libraries and stacks that teams can adapt quickly.

Exploring Radar SLAM: Advancing Localization and Mapping for Automotive, Robotics and Beyond

Reliable localization and mapping are foundational for autonomous vehicles and robots, yet camera-centric systems often degrade in fog, rain, snow, glare or low light—exactly when robust perception is most critical. In this talk, Amit Kumar, Director of Product Management and Marketing at Cadence and Amit Sulakhe, Director in the Vision Group at Cadence Pune, explore radar-based simultaneous localization and mapping (SLAM) as a complementary and, in some scenarios, alternative for building maps and estimating position under adverse conditions. They review the core principles of radar-based SLAM, what makes radar sensing fundamentally different from images and the practical challenges of working with sparse, noisy radar returns. They then examine recent advances in radar-only SLAM, including novel “iterative closest point” approaches for pose estimation and implementations optimized for real-time performance. Viewers will leave with a clear understanding of when radar SLAM outperforms vision-only methods, what accuracy is achievable today and which applications can benefit most from radar SLAM.

MAKING MODELS FIT THE EDGE

Small Language Models for Edge AI: Trade-Offs and Quantization in Practice

Large language models are powerful but often impractical for embedded and on-prem systems due to latency, cost, privacy and memory constraints. Small language models (SLMs)—typically comprised of single-digit billions of parameters or less—offer a deployable alternative but require different expectations and engineering choices. In this talk, Dwith Chenna, MTS Product Engineer, at AMD, introduces the SLM landscape and SLM applications together with performance and accuracy comparisons against LLMs. Dwith then examines the quantization techniques that matter for SLM deployment—gradient post-training quantization, SmoothQuant and activation-aware weight quantization—explaining how they work and how to compare them using metrics such as perplexity, task accuracy and runtime performance. Viewers will leave with a practical checklist for selecting, quantizing and evaluating SLMs for real edge systems.

Self-Compression for Edge Inference

Self-compression is a quantization-aware training technique to reduce neural network size and optimize performance for edge inference. By learning optimal bit depths for weights and activations during training, self-compression achieves significant reductions in memory footprint and bandwidth consumption while maintaining accuracy. The method employs high sparsity alongside low-bit representations, enabling efficient deployment on CPUs, GPUs, DSPs and NPUs without specialized hardware. Unlike traditional compression approaches, self-compression removes redundant weights and minimizes bits required for remaining parameters. Experiments demonstrate floating-point accuracy across applications including perception CNNs (as few as 3% of the original bits and 18% of weights retained) and LLMs (outperforming ternary compression in transformer-based language models). In this presentation, James Imber, Director of Research at Imagination Technologies, explains how self-compression works, its practical implementation and real-world benefits for embedded systems, and offers a simple yet powerful solution to reduce inference costs (execution time, power consumption, bandwidth and memory usage).

UPCOMING INDUSTRY EVENTS

 Multi-Agent & Hybrid AI with Intel AI Super Builder and OpenVINO Model Server

– Intel Webinar: August 26, 10:00 am PDT

See, Think, Act, Learn: Building Trustworthy Physical AI at the Edge

– Ambarella Webinar: September 2, 9:00 am PDT

SPIE AR, VR, MR Monthly Fireside Chat with FlexEnable

– SPIE Webinar: September 10, 9:00 am PDT

How Qualcomm Is Making Computer Vision Accessible Across Edge Verticals

– Qualcomm Webinar: September 17, 9:00 am PDT

Always-On Edge Perception Via a Heterogeneous Near-Memory AI Architecture

– FotoNation Webinar: September 22, 9:00 am PDT

Efficient Computer Vision at the Far Edge: Design and Training Under Constraints

– Lattice Semiconductor Webinar: September 24, 9:00 am PDT

Automotive Perception in the Physical AI Era: Imaging, Sensors and Computing

– Yole Group Webinar: September 29, 9:00 am PDT

Embedded Vision Summit: February 2-4, 2027, San Francisco, California

FEATURED NEWS

Intel has released OpenVINO 2026.3 with expanded model support

Perforated AI has released an enhanced ResNet-18 which outperforms ResNet-34, yet remains comparable in size and latency to ResNet-18

NVIDIA has released Alpamayo 2 Super, a frontier open model for robotaxis and autonomous vehicles, available for commercial use

Andes Technology has launched its AndesAIRE AnDLA I370 v2.0 Deep Learning Accelerator IP and the AndesAIRE NN SDK v1.2.0, to support ViTs, VLMs and SLMs

Microchip has released Revision 2.0 of its PolarFire FPGA Ethernet Sensor Bridge for standardized sensor integration leverage NVIDIA Holoscan Sensor Bridge technology

More News

EDGE AI AND VISION PRODUCT OF THE YEAR WINNER SHOWCASE

SiMa.ai Modalix MLSoC SoM (Best Edge AI Computer or Board)

SiMa.ai’s Modalix MLSoC SoM has been awarded the 2026 Edge AI and Vision Product of the Year Award in the Edge AI Computers and Boards category. SiMa.ai’s Modalix SoM (System on Module) is a compact, production-ready circuit board designed to run advanced AI directly on edge devices. Modalix SoMs are ideal for companies building physical AI systems—such as robotics, industrial automation, and intelligent vision applications—that need reliable, high-performance AI in real-world environments. Physical AI deployments often struggle with high power consumption, thermal limits, and expensive hardware redesigns. Modalix allows customers to modernize their existing platforms quickly, bringing more powerful AI closer to where the data is created without overhauling their systems.

Modalix combines purpose-built machine learning acceleration, Arm compute, vision processing, and high-bandwidth I/O in a single module. Its GPU-compatible software framework support and pin- and form-factor compatibility let customers upgrade from legacy or GPU-based designs with minimal friction—delivering higher performance per watt and faster time-to-production than competing Physical AI platforms. By delivering up to 50 TOPS of AI compute in under 10 watts, Modalix enables generative AI, large language.

Please see here for more information on SiMa.ai’s Astra™ Modalix MLSoC SoM. The Edge AI and Vision Product of the Year Awards celebrate the innovation of the industry’s leading companies that are developing and enabling the next generation of edge AI and computer vision products. Winning a Product of the Year award recognizes a company’s leadership in edge AI and computer vision as evaluated by independent industry experts.