Hands-on Physical AI with Kevin Cloutier

Kevin Cloutier is the North American Lead for Physical AI at CapGemini, a global engineering consulting firm. He is a computer engineer who is passionate about taking complex computations to the edge. He joins Chris to discuss the transition from legacy, statically programmed factory robotics to modern, imitation-learning-based “Physical AI,” building and calibrating the affordable $200 3D-printed SO101 robot arm using the LeRobot open-source framework or Intel’s Physical AI Studio, and orchestrating complex robotic systems entirely on local edge hardware.

Timeline

• Chris welcomes Kevin Cloutier to discuss “Physical AI” and how it relates to robots and things that move in the physical world. (00:00:15)
• Meeting at Embedded World: They reminisce about meeting at the Canonical booth in front of a robot demonstration, which you can see in the Embedded World Demonstration Video. (00:02:10)
• Trade Shows and Geography: Chris and Kevin discuss Embedded World in Nuremberg, Germany, upcoming SPS, and the high concentration of manufacturing robots in Europe compared to the US. (00:03:50)
• The “Unicorn Developer” Shift: Kevin talks about his background in computer engineering and how the “unicorn developer” has shifted from the full-stack web developer of the 1990s to someone who can operate, program, repair, and train physical robots. (00:06:05)
• Factory Robots vs. Edge Cases: Contrasting legacy, statically programmed factory robots (like ABB or Universal Robots hanging car doors) with modern generative AI approaches capable of handling real-world edge cases. (00:09:15)
• The Robotics Software Stack: Demystifying the layers above motor drivers, including Real-Time Operating Systems (RTOS), Linux, and message-passing frameworks like ROS (Robot Operating System) and ROS 2. (00:12:40)
• Orchestrating Systems of Systems: Kevin describes playing tic-tac-toe using Vision Language Action (VLA) models, where a higher-level camera and computer vision system orchestrate the coordinates for the movement model. (00:15:10)
• The Evolution of Compute: How modern silicon, integrated GPUs, and SOCs have allowed the massive, heavy control boxes of legacy robots to shrink down to a small NUC-sized device mounted directly on the robot. (00:18:05)
• The SO101 Robot Arm: Introducing the SO101 3D-printed robot arm from Hugging Face, which democratizes robotics by allowing anyone to build a leader-follower setup for around $200. (00:21:20)
• How “Backyard Engineers” Learn: Kevin’s advice for firmware and hardware engineers stepping into robotics: follow the documentation, get it running, and then ask questions about what you don’t know. (00:23:55)
• Calibration and the LeRobot Framework: A look at using the LeRobot open-source framework to calibrate hobby-grade motors and define their movement limits. (00:26:40)
• Recording “Episodes” via Imitation Learning: How users physically guide the leader arm to control the follower arm while a webcam records the visual and servo coordinate data. (00:28:50)
• Training the Model: Organizing data into short 5-episode chunks to make deletion easier, and training the model locally on NVIDIA GPUs or in the cloud. (00:31:30)
• From Training to Evaluation: Moving from training to evaluating the custom model, and asking questions about action chunking, model stutter, and operating frequency. (00:35:10)
• Sensor Fusion vs. Pure Vision: The current dominance of cameras in physical AI, and the potential to fuse accelerometers and time-of-flight sensors on mobile robots like Autonomous Mobile Robots (AMRs). (00:41:00)
• Real-World Calibration Challenges: Kevin shares a story of someone knocking over his camera boom at Hannover Messe and how he used April Tags to quickly recalibrate the camera’s physical coordinates. (00:43:15)
• The Reality of Humanoids: Debunking humanoid hype and explaining why full humanoids are further out than the public thinks, due to immense hardware cost, degrees of freedom, and safety. (00:45:50)
• The Puppeteer behind the Curtain: Kevin points out that many impressive humanoid demos, like the Unitree G1 at Hannover Messe, are actually being teleoperated by an engineer standing nearby. (00:48:40)
• The Complexity of Robot Subsystems: Using Steve Crunch and the book “Exploding the Phone” as an analogy for how modern robotic systems have become too complex for a single human to fully understand. (00:52:15)
• Understanding SmolVLA and SmolVLM: Diving under the hood of Vision Language Action models, which are fine-tuned transformer models that translate visual inputs into physical robotic coordinates. (00:55:10)
• Local AI and the NPU: How modern System-on-Chips (SOCs) let engineers run models locally on the CPU, GPU, or Neural Processing Unit (NPU) using optimization toolkits like Intel OpenVINO. (00:58:15)
• The Joy of the Physical World: Why working with physical hardware and robots is far more creative and rewarding than optimizing spreadsheets or SaaS applications. (01:04:30)
• “If the Robot Can’t Kill You, It’s Not Fun”: Kevin shares his colleague’s favorite metric for a truly exciting robotics project. (01:07:10)
• Advice on ROS and “Cobbling”: Starting with duct tape and bubble gum, and learning message-passing frameworks like ROS only when you need to coordinate multiple independent robots. (01:09:30)
• Physical AI Studio Demo: An invitation to see Kevin showcase Intel’s Physical AI Studio with multiple active robots at the upcoming AI Infra conference in Santa Clara. (01:14:45)
• Find Kevin online at his website cloutier.engineer or on LinkedIn. (01:20:10)