Sachin Kumar

Making Sim2Real smoother for roboticists. [GitHub link in post ⬇️] This minimal toolkit lets you train a reinforcement learning policy in Isaac Lab and deploy it directly on the Kinova Gen3 arm using only ROS 2. No simulation tools needed at runtime. ✅ Train with Isaac Lab using rsl_rl or rl_games ✅ Deploy in the real world using a lightweight ROS 2 setup ✅ Pretrained policies included so you can skip training ✅ Fully open-source and modular for custom tasks and robots ✅ Clean step-by-step guide for both fake and real hardware If you’re tackling Sim2Real, this toolkit makes it easier to go from a trained model to a moving robot, but without the mess. https://lnkd.in/diSAtGEy And thank you, Louis Le Lay, for sharing!

290

6 Comments

🦾 ROS developer Yifei Cheng of Remedy Robotics has released, "a generic ROS 2 interface for the Hugging Face LeRobot framework. It acts as a lightweight wrapper to connect any ros2_control or MoveIt compatible robot arm with the LeRobot ecosystem. A gamepad teleoperator for 6-DoF end-effector control and a keyboard teleoperator for joint position control is also provided." Full details on Open Robotics Discourse: 🔗 https://lnkd.in/gw66YeCB

252

8 Comments

With TinyML making machine learning accessible on microcontrollers, more embedded systems can now analyze data at the edge—detecting patterns, anomalies, and even making predictions without a cloud connection. 🚀   But some engineers argue that classical DSP algorithms (like FFT, Kalman filters, and wavelet transforms) are still more efficient for many low-power applications. So, which approach makes the most sense? 🤔   🔍 Why TinyML is Gaining Traction   ✅ Handles Complex, Nonlinear Data – ML models learn patterns that traditional DSP might struggle to define explicitly. ✅ Adaptability – Neural networks can be trained for specific environments, making them useful for fault detection, sensor fusion, and speech recognition. ✅ Cloud-Free AI – Processing on-device improves privacy, security, and real-time responsiveness. ✅ Better Hardware Support – MCUs with accelerators for AI inference (like Arm Ethos-U, Syntiant NDP) are making ML models more efficient.   ⚙️ Why DSP is Still a Strong Contender   ❌ Lower Computational Overhead – Classical DSP algorithms require far fewer operations than running a neural network. ❌ More Predictable Performance – DSP solutions have deterministic execution and low latency, critical for real-time control systems. ❌ Proven Reliability – Used in industrial, medical, and automotive systems for decades, DSP algorithms are trusted and well-optimized. ❌ Power Efficiency – TinyML requires neural network inference, which consumes more power than a simple filter or FFT in many cases.   💡 The Best of Both Worlds?   Some applications combine both approaches—using TinyML for pattern recognition and DSP for feature extraction and signal processing. For example: 🔹 Vibration analysis in predictive maintenance might use FFT for frequency extraction + TinyML for fault classification. 🔹 Speech recognition could rely on MFCC (DSP) for feature extraction + TinyML for classification.   📢 What’s Your Take?   👉 Is TinyML the future of low-power embedded intelligence? 👉 Are DSP algorithms still the best approach for efficiency and reliability? 👉 Have you used a hybrid approach?   Drop your insights below! 💬👇   #TinyML #DSP #EmbeddedSystems #MachineLearning #EdgeAI #IoT #LowPowerDesign #TechDebate #AIAtTheEdge #SignalProcessing

26

1 Comment

Diffusion policies forget the past. This method teaches robots to remember… with 3x better performance ⬇️ Robots need memory to act reliably over time, but adding history to diffusion-based policies usually makes performance worse and training more expensive. This new work introduces PTP, an auxiliary loss that brings back context understanding and makes long-history learning fast and effective. Why it matters ✅ PTP helps diffusion policies learn meaningful past-future connections ✅ Enables efficient training with cached short-context embeddings ✅ Boosts performance by 3x while cutting compute costs by over 10x ✅ Adds a test-time scaling trick that checks if the robot is paying attention to its own history Learning long-context robot policies just got way more practical. If we want smarter, more capable robots, this is a big step in the right direction. Thank you so much for sharing this, Marcel Torne Villasevil 🙏 Check out the paper and real-world tasks solved with PTP: Paper: arxiv.org/abs/2505.09561 Website: long-context-dp.github.io Code: https://lnkd.in/dPK2uG9x

99

2 Comments

🚀 Real-Time Metric Depth Estimation Real-time metric depth estimation with a ROS 2 node leveraging Depth Anything V3 and TensorRT acceleration. ✨ Key capabilities include point cloud generation from metric depth images, debug visualization with customizable colormap options, and support for configurable precision to balance performance and accuracy. 🤖 Use cases: ✅ SLAM and 3D mapping ✅ Augmented reality applications ✅ Robotic grasping and manipulation ✅ Scene understanding in dynamic environments ✅ Autonomous navigation and obstacle avoidance Awesome work by: Till Beemelmanns Stay tuned for more exciting developments and breakthroughs on the horizon! ✨ WISERLI YOLOvX NVIDIA OpenCV Roboflow Ultralytics Dr. Chandrakant Bothe Rohan Gupta Vishnu Mate Mohit Raj Sinha Tushar Gupta Prateeksha Tripathy Sinem Çelik Sharda Jadhav Neetu Shaw Shreya Nikam Siddhant Arora Mayuresh More Aaditya Bharti Pushpak Agrawal Anu Bothe Saurabh Tople Glenn Jocher Harpreet Sahota 🥑 Piotr Skalski Brad Dwyer Joseph Nelson Dragos Stan Arnaud Bastide Timothy Goebel Bharath kumar #Robotics #ComputerVision #DepthEstimation #ROS2 #YOLOvX #AI #AutonomousSystems

215

4 Comments

Which visualization tool choose for your Robotics project? Visualization is key in robotics, from debugging perception pipelines to analyzing recorded sensor and localization data. Ekaterina Marova has published on ReductStore blog a great article comparing three of the main visualization tool used in Robotics: ✅ RViz2: the ROS 2 native classic, perfect for live robot debugging and TF/URDF visualization. ✅ Foxglove: modern, collaborative, and browser-ready; great for synchronized playback and analysis. ✅ Rerun: lightweight, high-performance, and code-driven; ideal for large datasets and ML-driven robotics. Whether you’re working on navigation, SLAM, or perception, this deep dive will help you pick the right tool (or combination). 🔗 Read the full comparison and pick the best for your Robotics projects! Which visualization software are you using for your Robotics projects? Let's connect and share Robotics tips 🔽 #Robotics #ROS2

315

21 Comments

⚙️ Hardware-in-the-loop (HiL) testing automates and streamlines system development for robotics companies. It detects errors at an earlier stage, limiting the need for expensive manual tests or time-consuming test deployments. The entire process accelerates traditional robotics development, both in product development and maintenance. It is also essential when a system is deployed on highly customizable hardware, such as FPGAs, DPUs, TPUs, or integrated GPUs (to name a few). Fundamentally, it provides extra flexibility, supporting unusual hardware solutions. Another benefit of HiL in comparison to SiL (Software-in-the-Loop) testing is the wider coverage of the test. HiL testing can challenge lower layers of the robotics system (e.g., networking or device drivers). Our Digital Twin Simulation Platform RoSi has a dedicated feature for hardware-in-the-loop simulation. Book your demo to explore the benefits for your robotics project: https://lnkd.in/dXJPRyv7

27

CppCon, The C++ Conference 2024 - Chip Hogg: "Making Hard C++ Tests Easy: A Case Study From the Motion Planning Domain" youtu.be/8D7vpR9WCtw If you've ever struggled to write tests for domain-specific functions with complicated, real-world inputs, this talk is for you!  We'll use the Motion Planning component in a self-driving stack as a case study (although you won't need any prior Motion Planning experience to follow the talk).  In building objects for our test inputs, we faced the classic dilemma. If we make the objects simple, they're hardly meaningful, and the tests amount to little more than smoke tests.  If we try constructing more realistic objects, it takes tremendous amounts of boilerplate code (which obscures what is actually being tested) --- and what's worse, these finicky construction methods go stale and break easily as implementation details evolve. ... Full Abstract: sched.co/1gZh3 --- Chip Hogg is a Staff Software Engineer on the Motion Planning Team at Aurora Innovation, the self-driving vehicle company that is developing the Aurora Driver. After obtaining his PhD in Physics from Carnegie Mellon in 2010, he was a postdoctoral researcher and then staff scientist at the National Institute of Standards and Technology (NIST), doing Bayesian data analysis. He joined Google in 2012 as a software engineer, leaving in 2016 to work on autonomous vehicles at Uber's Advanced Technologies Group, where he stayed until their acquisition by Aurora in 2021. Chip's main role at Aurora is to design libraries that make it as easy as possible to write integration tests for Aurora's motion planner. He is also the lead author and maintainer of Aurora's units library, Au, and led its open-source release in 2023.

23

DeepSeek 🐳 just put a Linear Programming solver inside the GPU training loop. And it runs in 100 microseconds. If you want to know why the industry calls them "cracked engineers," look no further than their latest repo: LPLB (Linear-Programming-Based Load Balancer). Here is the Voodoo: In MoE training, load balancing is usually done with greedy heuristics. Why? Because you can't afford to think. The GPU is hungry, and you have milliseconds to decide which token goes to which expert. DeepSeek looked at "greedy heuristics" and said, "No. We want the mathematically optimal solution every single batch." So they built a custom Linear Programming solver that: - Runs on bare metal: It bypasses the CPU entirely. - Is surgical: It runs on a single Streaming Multiprocessor (SM) using an Interior Point Method. - Is fast: It solves the optimization problem in ~100µs. While the other 100+ SMs are crunching matrices, this one lonely SM is solving a complex logistics optimization problem in real-time to ensure perfect load balancing for the next micro-batch. This isn't just "optimizing code." This is: - Using cuSolverDx to run linear algebra in registers. - Treating a distributed systems problem as a math problem. - Refusing to accept "good enough" routing. This is what happens when you let math nerds write low-level GPU kernel. Check the repo. It’s startlingly elegant. https://lnkd.in/gy79x3hN

225

1 Comment

💡Use Example Models to quickly learn specific features of CODE V This example discusses a workflow on how to remove an element from a starting point spherical design by introducing aspheres. The workflow is demonstrated with a mid-wave infrared (MWIR) system, although the technique is not limited to MWIR. With the example models, you can explore CODE V’s extensive design and analysis capabilities as you work on practical examples. Engineers can quickly learn how to use a feature and be instantly more productive. Get a head start designing your next imaging application with the example models in CODE V. #ExampleModelMondays #codev #codevfeatures #optics #opticaldesign #opticalengineering #designbrilliance

36

Precision Foot Sensing for Humanoid Robotics 🤖 Effective locomotion in humanoid robots requires accurate measurement of ground reaction forces to compute center-of-mass (CoM) shifts during gait. A recent implementation features six strain gage-based load cells integrated into each foot, each with a 200 N capacity. This configuration enables multi-axis force resolution—supporting real-time analysis of vertical load, shear, and moment forces throughout the walking cycle. The resulting data improves stability control, posture adjustment, and motion planning. By capturing subtle force vectors at the foot-structure interface, this approach supports more adaptive and human-like movement across variable terrain. Strain gage technology continues to provide the sensitivity and bandwidth needed for advanced robotic applications.

66

14 Comments

Developing with ROS2? Speed up your robotics journey with Torizon + Toradex SoMs - a powerful combo that simplifies integration, boosts maintainability, and accelerates deployment. In this article, discover how Torizon streamlines ROS2 development so you can: 🔹 Prototype faster 🔹 Integrate ROS with ease - no headaches with Yocto 🔹 Easily machine-to-machine communicate 🔹 Focus on Application Development Get to market faster with a robust platform built for modern robotics. 📖 Read the full article: https://lnkd.in/dEZWyNBy #Robotis #EmbeddedSystems #Linux #EdgeComputing #IoT #DevOps #Torizon

10

Creating robot paths shouldn’t mean hours of manual work and constant readjustments.   With Visual Components 5.0, our new automatic path solver makes programming faster and more reliable, even in complex cells and tight spaces.   And that’s just one of the updates. From automated programming workflows to our new Fronius Welding connector, we’ll be showing everything new in 5.0 at SCHWEISSEN & SCHNEIDEN.   Find us at MESSE ESSEN GmbH, Hall 3, Booth 3B12, 15.–19.9.2025   Here’s a quick preview from Jarkko Siltamäki, Head of our robotics business. #SchweissenUndSchneiden #RobotProgramming #OfflineProgramming #WeldingAutomation

77

3 Comments

Bridging the gap between legacy C++ models and modern simulation workflows? CxxWrap.jl makes it easy to integrate high-performance C++ code into Julia—enabling faster development with #JuliaSim. Reuse, extend, and accelerate your models today! https://lnkd.in/eED8aKA8 #JuliaLang #Simulation #Engineering #Interoperability

31