Axis Robotics has launched Axis Sim Dataset V1, a substantial open-source simulation dataset for Franka arm manipulation. This dataset comprises over 50,000 human-teleoperated trajectories across 207 distinct manipulation tasks and 60,000 scene variations. It has become the most downloaded open-source simulation dataset for Franka manipulation on Hugging Face, with over 160,000 downloads. Benchmarks show that pretraining with V1 significantly improved performance, surpassing a comparable baseline. Axis Robotics is developing a comprehensive data engine for Physical AI, integrating large-scale simulation, real-world capture, and humanoid locomotion. The company secured $12 million in seed funding to advance this ambitious project. Their philosophy challenges the notion of relying solely on "clean data," arguing that diverse and even noisy data, when aggregated, can lead to robust policies. The dataset includes a wide array of tasks like pick-and-place, stacking, and tool use, collected via their Axis Hub teleoperation platform. Continual pretraining on V1 demonstrated consistent performance improvements as the data scale increased, indicating the benefits of diversity and coverage. Axis Robotics emphasizes that the dataset is a product of a larger, compounding data engine that continuously learns from model performance and failures. This engine utilizes simulation, egocentric real-world capture, loco-manipulation, and human-gated DAgger post-training to gather data. Beyond open-sourcing, they collaborate with robot embodiment companies to create customized data pipelines and model priors, as exemplified by their work with Booster Robotics. Axis Robotics asserts that the future of Physical AI lies in dynamic data engines, not static datasets, promoting continuous progress through scale, diversity, and closed-loop learning.