Daniel Pelaez

The paper introduces SVII‑3D, a novel computer vision framework designed to automate the creation of detailed roadside infrastructure inventories using sparse street imagery. Traditional methods struggle with limited image coverage, poor robustness, and inaccurate localization, which hinders the scalability of smart city digital twins and asset lifecycle management. SVII‑3D addresses these challenges by first combining LoRA‑fine‑tuned open‑set object detection with a spatial‑attention matching network to reliably associate detected infrastructure elements across multiple sparse views. This association step improves robustness over typical feature matching and enhances correspondence between images taken from different perspectives. Beyond detection, the system incorporates a geometry‑guided refinement mechanism that corrects structural errors to produce decimeter‑level 3D localization of roadside assets. This high positional accuracy is crucial for precise digital twin models. Additionally, SVII‑3D integrates a Vision‑Language Model (VLM) agent with multi‑modal prompting to go beyond geometric mapping and automatically interpret the operational state and fine‑grained attributes of infrastructure elements. Experimental results in the paper show that this unified approach significantly enhances identification accuracy and minimizes localization errors, offering a scalable and cost‑effective solution for high‑fidelity infrastructure digitization even from sparse visual data. https://lnkd.in/g2ssbkTw

1

Evolving SkyRL into a Highly-Modular RL Framework https://lnkd.in/eiCEAdS6 SkyRL-v0.1: Making SkyRL Modular In the [original release of SkyRL](https://lnkd.in/eDQFxZef), we introduced an agentic layer in the RL stack for multi-turn tool use LLMs, optimized for long-horizon, real-environment tasks like SWE-Bench. Today, we are upgrading SkyRL to a highly-modular RL framework to train LLMs with the introduction of two key additions: 1) A **modular**, **performant** **RL framework** for training LLMs. SkyRL makes it easy to prototype new training algorithms, environments, and training execution plans — without compromising usability or speed. 2) A **gymnasium of tool-use tasks** with a simple environment interface and an ****out-of-the box library of popular tasks ****such as math, code, search, and SQL. SkyRL’s modularity enables easy implementation of real-world improvements—like async training, heterogeneous hardware, and new environments — with **under 100 LoC** and up to **1.8× faster training**.

1

iVue Robotics Founder, Jacob Andrews, returns to Denmark Drone Park with Lead Electrical Engineer, Gerardo Salinas, cracking jokes and reflecting on data privacy with LLMs while prepping the latest build. After clarifying some online prop diagrams are just plain wrong—urging viewers to "search ArduCopter specifically"— Jacob and Gerardo run into a group of engineers and PhDs from Georgia Tech who are out exploring the field. Between troubleshooting a pricey VTX connector and navigating a "different connector" on a Develop Air mini battery, Jacob nearly gets a clean bill of health before a "compass not healthy" message proves he "spoke too soon". Watch the full video here: China #1... For Now & 2 Fingers or 3 Fingers - Meeting Mohammad, Sloppy & Loose @ Denmark Drone Park https://lnkd.in/eENEeNWm Recorded on 11/16/25. iVue RoboControl (iVue RC) is available at worldsivue.com. Develop Air is available to buy or build DIY at ivuerobotics.com. #drone #flighttest #robotics #engineering #ArduCopter #GeorgiaTech #shorts

1