Flux 3 Action: A 7B open-weight world action model for robots (huggingface.co)

The newly announced FLUX 3 Action is a groundbreaking 7 billion parameter open-weight world action model (WAM) designed to enhance robotic performance in dynamic environments. Fine-tuned on the DROID dataset and amassing a 42.92% success rate on the RoboLab benchmark, FLUX 3 Action outperforms the 16B Cosmos 3 Nano model, which scored 36.8%. This model integrates checkpoints from both DROID and the SO-101 robotic arm into the LeRobot framework, making it easily trainable for various tasks, including pick-and-place operations, and offers adaptability for other applications like gaming and drone control. What sets FLUX 3 Action apart is its ability to jointly predict future frames and actions, utilizing a unique architecture that processes video and action tokens together. The model is trained using NVIDIA's GB200 systems and leverages custom kernels for fine-tuning and deployment on NVIDIA Jetson devices. Its versatility was further demonstrated through its performance in scripted games, where it learned optimal control policies in a fraction of the time it would take in real-world settings. By enabling cross-domain applications and providing an open-source framework, FLUX 3 Action significantly advances the capabilities of AI/ML models in robotics and beyond, fostering innovation in autonomous systems.