LeRobot v0.6.0: Imagine, Evaluate, Improve (huggingface.co)

LeRobot has launched version 0.6.0, featuring groundbreaking developments including world model policies that enable robots to imagine future scenarios, new vision-language-action (VLA) models, and an advanced reward models API. The significance of this update lies in its potential to enhance robotic decision-making through improved training methodologies and rewards, fostering advancements in areas like autonomous navigation and manipulation. Notable policies like VLA-JEPA utilize latent space predictions to foster imaginative training while ensuring no added cost during inference, and FastWAM challenges the necessity of test-time future imagining by operating without it at inference. The update introduces an array of simulation benchmarks and tools designed to streamline robot policy evaluation and deployment. Key technical advancements include a unified reward models API that allows for success detection from raw video, accelerated data loading for training processes, and enhanced language annotation capabilities for richer dataset management. Additionally, new features for seamless cloud training and deployment workflows enable users to integrate corrections into their learning cycles efficiently. Collectively, these enhancements position LeRobot as a vital hub for researchers and developers in the AI community, paving the way for more sophisticated robotic systems that leverage imagination and real-world feedback to improve performance.