Saeed Jolghazi, PhD

🚗 The Patterns We See That Autonomous Vehicles Don’t (Yet) Real-world autonomy doesn’t just need fast reaction times, It needs something subtler, the system’s internal understanding has to stay consistent with what’s actually happening in the environment. 🎥 The Setup I replicated a real highway clip as accurately as possible and ran one of the initial iterations of the L3 AV stack I built, which lacks contextual inference. (The scenario isn’t an exact pixel-level replication, it is designed purely to compare human pattern-based reasoning with the AV stack’s decision logic.) Video layout: Top: Raw dash-cam feed Bottom row (the AV stack’s POV): • Left: Perception overlay (what the system “sees”) • Center: maneuver visualization (What the AV decides, plans and reacts) • Right: Trajectory plot (ego + actors + ego's past/future path) 🧠 The Reasoning Difference • Human driver (intuition + pattern recognition): The left lane looked faster, but the driver sensed something subtle, a long chain of tightly packed vehicles , usually means a hidden slowdown ripple. So they held the center lane more open space, lower risk, fewer surprises. We subconsciously evaluate patterns, comfort, and “how this will play out”. (The reasoning described here is a hypothesis. Decisions and reasoning varies person to person, But patterns like traffic density, speed, other drivers behavior and other cars trajectory are things we evaluate subconsciously, We try to take calculated risks) • AV stack (object-by-object logic): The system treated each car as an isolated object, not a pattern. Left lane vehicles were moving slightly faster, so the AV predicted better progress and attempted a lane-change. Technically correct, but context-blind, no understanding of congestion dynamics. 💡 What This Shows This isn’t about blaming the AV or praising the human. It’s a glimpse into the gap between pattern recognition and true anticipation. We’re good at reading subtle patterns like traffic density, traffic rhythm, how “trapped” a lane feels. Most AV stacks are still very good at local optimization and very early at contextual understanding. ❓Question for You Which is harder for an AV to learn, predicting what others will do or understanding why they might do it? Should we incorporate psychological priors (driver styles, risk tolerance), or is that too unpredictable for safety? Curious to hear your take. #AutonomousVehicles #AV #ML #robotics #MATLAB #ComputerVison #OpenCV #Mapping #ADAS #roboticsengineering #autonomousdriving #AI #selfdriving #PathPlanning #IntelligentSystems #Technology

267

41 Comments

The Boundary Layer Effect on Heat Transfer: Insights from Research Did you know that the boundary layer in fluid dynamics significantly impacts heat transfer efficiency? 🔍 Recent research from MIT's Rohsenow Kendall Heat Transfer Laboratory reveals that heat transfer coefficients in laminar boundary layers are strongly sensitive to upstream flow history, including leading-edge configurations and wall temperature variations. As a boundary layer develops along a surface: 🔹 Heat transfer tends to decrease with increasing boundary layer thickness 🔹 The thermal boundary layer is typically slightly thicker than the velocity boundary layer for most fluids 🔹 For air (Pr ≈ 0.7), the thermal boundary layer is about 10% thicker than the velocity boundary layer Contrary to simplified models that assume an abrupt transition point, experimental data shows an extended transition region where turbulence gradually develops. This transition region can be as long as the laminar region itself. The heat transfer coefficient varies significantly across these regions: 🔹 Laminar region: Predictable, analytically solvable heat transfer 🔹 Transition region: Gradually increasing heat transfer rates 🔹 Turbulent region: Higher heat transfer rates due to enhanced mixing Turbulent boundary layers exhibit larger shear stress and higher heat transfer rates compared to laminar ones. This is why engineers often intentionally trigger turbulence in heat exchangers to improve thermal performance. For uniform heat flux vs. uniform wall temperature conditions: In laminar flow: Heat transfer coefficient is 36% higher for uniform heat flux In turbulent flow: The difference reduces to only about 4% The practical implications? Understanding boundary layer development is crucial for designing efficient thermal management systems in electronics cooling, power generation, and aerospace applications. #ThermalEngineering #HeatTransfer #FluidDynamics #BoundaryLayer #CFD #ANSYS #AltairCFD #BETACAE Citations: https://lnkd.in/g8597-VF https://lnkd.in/gMJjn4qq https://lnkd.in/gxMBtM_z https://lnkd.in/g62fxcgF https://lnkd.in/gkCub4hq https://lnkd.in/gw9VZuZA https://lnkd.in/gXJX3pX6 https://lnkd.in/g2citubh

10

👩‍💻 Are Zoox and Waymo Automotive Companies or Tech Companies? Five years ago, I made a significant shift in my career: moving from supporting some of the world’s most innovative high-tech customers to working in the automotive industry. It wasn’t just a change of industries— it was stepping into a space where hardware meets AI, where vehicle safety meets software reliability, and where two massive industries are colliding to redefine mobility. As I immersed myself in this new world, one question kept coming up — both for the companies leading the charge and for professionals like me: 🤔 Are companies like Zoox and Waymo automotive companies, tech companies, or something entirely new? Zoox, acquired by Amazon in 2020, is building a fully autonomous vehicle from the ground up — not retrofitting existing models. This places it firmly in the automotive realm for engineering, safety, and manufacturing. But its true edge is its software stack: perception, localization, motion planning, and fleet orchestration — capabilities traditionally owned by robotics and AI firms. Waymo, born from Google’s self-driving project, took a different approach. By partnering with OEMs like Jaguar and Chrysler, it integrates its autonomy platform into existing vehicles. This positions Waymo less as a manufacturer and more as a mobility technology platform, akin to how Android powers various smartphone brands. 🤔 Why Does this matter you might ask? This distinction influences how these companies operate: 👉 Capital Structure: Automotive requires heavy investment in factories and regulatory approval; tech prioritizes software, cloud, and data infrastructure. 👉 Business Models: Automakers sell vehicles; tech companies license platforms or charge per use. 👉 Regulation: Automotive focuses on crash safety and production standards; tech focuses on cybersecurity, AI ethics, and data compliance. 👉 Talent: Automotive attracts mechanical and operational expertise; tech attracts AI, software, and cloud talent. Zoox integrates both worlds — building vehicle + software. Waymo focuses on scaling autonomy software across partners. 🤔 The real answer may not be “automotive” or “tech.” Both companies are moving toward Mobility-as-a-Service — deploying autonomous fleets that could reshape urban transport. Ownership gives way to orchestration: fleets, logistics, and city-scale mobility infrastructure. The automotive and technology industries are no longer separate lanes — they’re converging into a new mobility ecosystem. Zoox and Waymo sit at the center of this transformation, challenging legacy categories. ✅ The question of automotive or tech? is certainly outdated. The answer might be: both — and something more. #AutonomousVehicles #MobilityInnovation #AutomotiveIndustry #MobilityTech #SoftwareDefinedVehicles #FutureOfTransportation #MobilityAsAService #InnovationStrategy

3

Most automotive companies chase the biggest markets they can find. That’s usually a mistake. Trillion-dollar markets look attractive from the outside until everyone else piles in, competition explodes while margins collapse, and you need insane scale just to survive. The Ford F-150 is a trillion-dollar market. But Ford can afford to hypervertically integrate everything because the volume justifies the investment. On the other end of the spectrum? $5B markets. But automakers can’t manage something that small. The hypervertical integration isn’t justifiable and you can’t build the manufacturing infrastructure to make the economics work. The real sweet spot sits in the middle. ~$50B markets. Small enough that big OEMs won’t justify the investment in ground-up platforms and vertical integration. Large enough to support full control of the supply chain and generate meaningful revenue. Not because the tech is better, but because no one else can afford to do it. In these markets, we build with manufacturing capacity 1–2 orders of magnitude cheaper than industry norms and because of that we can command 50% margins. That’s semiconductor margins. Not automotive margins. The advantage isn’t innovation for innovation’s sake. It’s choosing markets where discipline beats scale. That’s how you compete with companies 10x your size and win.

222

7 Comments

Here's our latest paper with a deeper analysis of Waymo’s performance across 11 different crash types compared to human drivers. It also offers new insights into Waymo’s positive impact on serious injury crash rates. Kristofer Kusano John Scanlon, Ph.D. Tilia Gode Tim McMurry Yin-Hsiu (Aaron) Chen blog: https://lnkd.in/d4wVcNqE paper: https://lnkd.in/d6pfhihi

185

7 Comments

BACK-TO-BASICS: A young engineer, which we'll name Jean Claude, is asked to compute the vertical deflection of a notched beam (as shown below to scale). He proceeds to set up a finite element model consisting of three beam elements (Timoshenko beam theory), each element has a depth and width corresponding to the physical dimensions of the beam for each corresponding segment. If you were Jean Claude's supervisor, what would you say about this model? Does it overestimate the deflection, or does it underestimate it? Explain. The material is homogeneous, isotropic, linear, and elastic.

81

24 Comments