ACT-Lab Visual Pipeline: From Demonstrations to Real-World Skills

🚀 ACT-Lab Visual Pipeline: from demonstration capture to model validation I’ve been working on a new visual pipeline for ACT-Lab, designed to bring the complete robot-learning workflow into a single interface. The idea is simple: make it easier to go from data collection to a validated ACT model without constantly switching between scripts, terminals, and separate tools. With the current version of the pipeline, the workflow covers: Camera setup and ArUco verification Dataset capture from robot demonstrations Multimodal data recording, including images, robot states and actions ACT model training Training monitoring and checkpoint management Model validation and execution The goal is to make ACT-Lab more modular, easier to use, and more suitable for experimenting with different robots, datasets, tasks, and learning strategies. This is still an early version, but it is already becoming the central interface for the capture, training, and validation process in ACT-Lab. As future work, I would like to expand the framework with Intel RealSense cameras, support for additional robotic platforms, more sensors, improved multimodal data handling, new perception modules, and other interesting capabilities that can make ACT-Lab even more flexible for research in Physical AI and Embodied AI. There are still many ideas ahead, but step by step the platform keeps growing. One step closer to the first public release of ACT-Lab. 🤖 From demonstrations to real-world skills. #ACTLab #PhysicalAI #EmbodiedAI #RobotLearning #ImitationLearning #Robotics #ACT #MachineLearning #ComputerVision #RealSense #ArtificialIntelligence #OpenSource #RobotManipulation