Post by M A HAFIZ

I’m excited to share a small demo from my current humanoid robotics work. In this experiment, I controlled an ARMAR-6 humanoid robot in Isaac Sim using webcam-based human pose imitation. The system detects my upper-body and arm motion with MediaPipe, classifies the intended pose using a lightweight ML model, and sends live commands to the simulated robot. ✅ The robot can respond to different right-arm poses, including: ✅ Natural arm-down position ✅ Front/up open-arm pose ✅ Right-side and left-side arm movement ✅ Elbow-close motion ✅ Hand/finger open-close gestures One key challenge was separating the front/up open-arm pose from the elbow-close motion, because both can look very similar from a single 2D webcam view. To solve this, I recorded pose samples for different motion classes and trained a lightweight classifier to map webcam landmarks into robot control commands. This is still an early prototype, but it shows a practical direction toward intuitive robot teaching and teleoperation using simple human demonstrations. ✅ Tools used: Isaac Sim, ARMAR-6, Python, MediaPipe, scikit-learn, webcam-based pose estimation. ✅ Next step: improving robustness, adding smoother trajectory generation, and connecting this idea with imitation learning for task-level robot behavior. #Robotics #HumanoidRobotics #IsaacSim #RobotLearning #HumanRobotInteraction #ImitationLearning #Teleoperation #ComputerVision #MachineLearning #Python #MediaPipe #Simulation #EmbodiedAI #AI

Video Content