Black Forest Labs debuts FLUX 3 Action, tops RoboLab leaderboard

NVIDIA: NVIDIA develops robotics hardware, software frameworks, and benchmarks including the RoboLab simulation environment for evaluating general-purpose robotic policies. It also creates competing models such as Cosmos and GR00T for vision-language-action and world action tasks. In this context, NVIDIA's RoboLab benchmark and models serve as the primary comparison points for Black Forest Labs' new release. RoboLab: RoboLab is an NVIDIA-developed simulation benchmark designed to test robotic policies across 120 tasks involving visual understanding, relational reasoning, and procedural skills at varying difficulty levels. It aims to provide a more robust evaluation than older saturated benchmarks. Black Forest Labs' FLUX 3 Action claims the top position on the RoboLab-120 leaderboard according to the company's reported results. FLUX 3 Action: FLUX 3 Action is a 7-billion-parameter open-weights World Action Model developed by Black Forest Labs for translating camera observations, robot state, and natural language instructions into physical actions. It extends the company's FLUX 3 family, which was first previewed in July with Action initially limited to select partners. The model is designed for practical deployment in robotics, simulations, games, and computer use, with plans for full open release including code and fine-tuning support. Black Forest Labs: Black Forest Labs is a German AI startup founded by researchers behind latent diffusion and Stable Diffusion technologies. It initially gained recognition for its FLUX family of image and video generation models and has expanded into robotics applications. In this news, the company is debuting FLUX 3 Action, an open-weights robotics model that builds on its video pretraining to enable physical action prediction and control. Open Robotics Ecosystem: Multiple organizations including NVIDIA and Ai2 have released open or downloadable robotics models with tools for local fine-tuning and adaptation to specific hardware. World Model Pretraining: Approaches that pretrain models on video to learn motion and cause-and-effect representations are being adapted across robotics teams to reduce reliance on large task-specific datasets. Deployment Considerations: Inference speed and model size are key factors for real-world robot policies, as they affect reaction times and feasibility on varied hardware setups.