How machines will learn to see before they think in the Physical AI era

AI has advanced at an incredible pace, but why do our machines still struggle with basic spatial awareness? In this new article for EE Times, we argue that the biggest bottleneck to the next era of automation is perception, not intelligence. To unlock the future of automation and its incredible value, we must first solve the sensing problem - at scale. I discuss how Silicon Photonics is making it possible to "democratize autonomy" by giving machines the power to see. Check it out: https://lnkd.in/epaCy4hm #PhysicalAI #LiDAR #Sensing #Autonomy #SiliconPhotonics

I'm thrilled to announce that my paper has been accepted at the top-tier AI conference AAAI-26, titled “PEFT-DML: Parameter-Efficient Fine-Tuning Deep Metric Learning for Robust Multi-Modal 3D Object Detection in Autonomous Driving.” The proposed method unifies multi-modal sensor data (LiDAR, radar, camera, IMU, GNSS) into a shared latent space where each modality’s features are projected, fused with cross-attention and gating, and used for 3D object detection even when some sensors fail. #MachineLearning #DeepLearning #ComputerVision #AutonomousDriving  #MultimodalAI #3DObjectDetection #LoRA #ParameterEfficientTuning #PEFT #DeepMetricLearning #SensorFusion #AIResearch #PyTorch #AIEngineer 

Today, I implemented Graph-based SLAM using slam_toolbox in ROS2....... here’s a live view from RViz2. The robot is actively mapping its environment using LaserScan data (red points). while generating an occupancy grid map (gray cells) in real time. The blue trajectory represents the robot’s optimized path.... refined continuously through pose graph optimization for consistent and accurate mapping. Watching the map evolve as the robot explores the environment never gets old....... it’s the perfect blend of perception, localization, and intelligence working together. #Robotics #ROS2 #SlamToolbox #GraphSLAM #Mapping #LiDAR #AutonomousRobots #Navigation #AI #RoboticsEngineering

MIT just leaked their full Drones + Computer Vision curriculum online. (And almost nobody’s talking about it.) Save this post. Right now. Not “later.” Not “when you have time.” Now. Because the startup that beats you won’t have more funding. They’ll just know the right systems. Here’s the roadmap that actually builds real drone AI skills 👇 ⸻ 1️⃣ START HERE → MIT 6.034: Artificial Intelligence The foundation everything else builds on. Search, reasoning, perception — everything drones need to think. 🔗 https://lnkd.in/dc3Qa4iR 🔗 YouTube lectures: https://lnkd.in/dHzW5YeD ⸻ 2️⃣ THEN THIS → ETH Zurich: Vision for Mobile Robotics This is where drones actually learn to see. Multiple view geometry, visual odometry, SLAM, event cameras — the works. 🔗 https://lnkd.in/dMuGnzx7 ⸻ 3️⃣ PICK YOUR WEAPON ▸ Perception Track → Stanford CS231n: CNNs for Visual Recognition Transformers, detection, segmentation — all the neural eyes your drone needs. 🔗 https://lnkd.in/deYY2BTM ▸ Control Track → UPenn Robotics: Perception (Coursera) The math behind 3D reconstruction and navigation. 🔗 https://lnkd.in/dNQ7yasM ⸻ 4️⃣ BONUS ROUND (when you’re ready for the deep stuff) ▸ Visual Odometry Tutorial by Scaramuzza 🔗 https://lnkd.in/dnE8FHBx ▸ UZH Research on Visual-Inertial Odometry 🔗 https://lnkd.in/dw_XNeRt ⸻ While you’re watching drone reviews, someone’s finishing these five courses and joining Skydio or Anduril. Total cost: $0 Total time: 180 days ROI: Drone Vision Engineer or Robotics AI role ♻️ Repost to save someone £50K and 2 years of confusion. #AI #Drones #UAV #Robotics #ComputerVision #SLAM #MIT #Stanford #DeepLearning #Autonomy #London

Autonomous Navigation Using Machine Learning Autonomous navigation means a robot, drone, or vehicle can safely go from A to B by itself. Machine learning gives these systems the ability to perceive complex scenes, make decisions under uncertainty, and adapt to new environments in real time. Perception and Sensor Fusion What it does: turn raw sensor streams (camera, lidar, radar, IMU) into actionable scene understanding: objects, drivable space, and dynamic agents. How ML helps: deep CNNs and transformers detect and segment objects; learned sensor fusion produces robust, redundant state estimates for navigation. Localization and Mapping What it does: estimate the agent’s pose and build or use maps for planning. How ML helps: learned visual–inertial odometry and data-driven SLAM components improve robustness in challenging conditions and reduce reliance on perfect sensors. Planning and Decision Making What it does: choose routes and generate collision-free trajectories that respect dynamics and rules. How ML helps: reinforcement learning and imitation learning teach complex behaviors; learned policies can be combined with classical planners for long-horizon, safe decisions. Control and Execution What it does: convert planned trajectories into steering, throttle, and actuation while handling real-world dynamics. How ML helps: learned controllers and residual models adapt to model mismatch and improve tracking performance under real disturbances. Safety, Sim To Real, and Data What it does: ensure reliable operation under uncertainty and edge cases. How ML helps: large-scale simulation, domain randomization, and synthetic data enable safe training; uncertainty estimation and runtime monitors reduce deployment risk. Quick student action plan Start in simulation with ROS and Gazebo or PyBullet. Build a pipeline: sensor input → perception model → localization → planner → controller. Use domain randomization and test edge cases before hardware trials. #AutonomousNavigation #MachineLearning #Robotics #SelfDriving #SLAM #SensorFusion #ReinforcementLearning #SimToReal #ControlSystems #AIForGood

Technology hidden inside dragon fly . 🧠 1. Neural Processing Power Dragonflies have tiny yet powerful brains—they can process visual information about 200 times faster than humans. Their neural circuits predict prey movement, allowing them to intercept other insects mid-air with over 95% hunting accuracy — that’s better than most guided missiles! Scientists study these circuits to develop autonomous drones and AI tracking systems that mimic this precision. --- 👁️ 2. Compound Eyes (Advanced Vision System) Dragonflies have ~30,000 ommatidia (facets) per eye, giving them nearly 360° vision. They can detect ultraviolet, visible, and polarized light, making them extremely sensitive to motion. This inspires 360° cameras, surveillance optics, and vision sensors for drones and self-driving vehicles. --- 🦾 3. Wing Structure and Flight Mechanics Each of the four wings can move independently, allowing hovering, backward flight, and rapid turns. The wing veins are natural shock absorbers and energy redistributors — an example of biomimetic engineering. Engineers study this for flapping-wing drones (ornithopters) that can maneuver in tight spaces. --- 🔋 4. Energy Efficiency and Muscular Control Dragonflies are among the most energy-efficient flyers. They use direct flight muscles attached to each wing for fine control, unlike most insects that use indirect muscles. Their design has inspired research in micro air vehicles (MAVs) powered by lightweight actuators. --- 🧬 5. Biological Sensors Tiny hairs on their bodies detect wind direction and speed. Mechanoreceptors on their wings sense strain and deformation. This natural sensory feedback system works like a built-in flight computer, adjusting their wingbeat in real time. --- 🛰️ 6. Military and Technological Inspiration Defense researchers have studied dragonflies for bio-inspired drones and micro surveillance systems (sometimes called “micro UAVs”). For example, the DragonflEye Project (by Draper & Howard Hughes Medical Institute) created a cyborg dragonfly by integrating a miniature steering system into a real insect’s nervous system for controlled flight.

Microbiologists study dragonflies for two primary reasons: their bactericidal wings and their complex gut microbiomes. Dragonflies serve as valuable models for both antimicrobial research and understanding insect-microbe relationships. Nanopillars: Dragonfly wings are covered in tiny, spiky structures called nanopillars (or nanostructures). These are mechanical features—not chemical agents—that physically interact with bacteria. https://lnkd.in/dDkrrm6g

Came across beautiful yet interesting.....!! Projects like "Dragonfly"beautifully merge biology with robotics, transforming real dragonflies into living, steerable micro-drones. By integrating tiny navigation systems directly into the insect’s nervous network, scientists have created a seamless blend of nature and technology, a breakthrough that redefines what’s possible in bio-inspired engineering and autonomous flight. #DragonflEye #AI #Drones #FutureTech #NatureInspired #NeuroEngineering #ScienceAndTechnology

AI-powered Drone for Transmission Line Fault Detection ⚡🤖 I'm excited to share my latest project — a smart drone system designed to detect faults in power transmission linesusing Artificial Intelligence and Computer Vision. The drone autonomously scans powerline routes, captures real-time images, and uses a deep learning model to identify potential issues like damaged insulators, broken lines, or vegetation interference — even in areas that are hard to reach by vehicles such as mountainous regions. This project merges my passion for embedded systems, AI, and drone technology to enhance safety, reduce inspection costs, and improve powerline maintenance efficiency. A big step toward smart energy infrastructure 🔋🌍 #AI #DroneTechnology #ComputerVision #SmartEnergy #PowerlineMonitoring #Innovation #EmbeddedSystems

AI-powered Drone for Transmission Line Fault Detection ⚡🤖 I'm excited to share my latest project — a smart drone system designed to detect faults in power transmission linesusing Artificial Intelligence and Computer Vision. The drone autonomously scans powerline routes, captures real-time images, and uses a deep learning model to identify potential issues like damaged insulators, broken lines, or vegetation interference — even in areas that are hard to reach by vehicles such as mountainous regions. This project merges my passion for embedded systems, AI, and drone technology to enhance safety, reduce inspection costs, and improve powerline maintenance efficiency. A big step toward smart energy infrastructure 🔋🌍 #AI #DroneTechnology #ComputerVision #SmartEnergy #PowerlineMonitoring #Innovation #EmbeddedSystems