Axis Robotics has recently unveiled Axis Sim Dataset V1, an expansive open-source simulation dataset designed for Franka arm manipulation. This dataset includes over 50,000 human-teleoperated simulation trajectories covering 207 manipulation tasks and over 60,000 scene variants using a simulated Franka Research 3 arm. The full dataset, training code, and benchmarks are all accessible to the public.
The release of V1 garnered significant attention, with over 160,000 downloads, making it the most downloaded open-source simulation Franka manipulation dataset on Hugging Face. In benchmark tests, continual pretraining on V1 resulted in a significant improvement, surpassing a volume-matched RoboCasa baseline, with transparent and verifiable results.
Axis Robotics is on a mission to build an extensive data engine for Physical AI, encompassing large-scale simulation, real-world capture, humanoid loco-manipulation, and human-gated DAgger post-training. The company recently secured $12 million in seed funding led by Hack VC, with additional contributions from Nomad Capital, Pi Network Ventures, 10K Ventures, and angel investors.
The foundation of Axis’s approach challenges the conventional belief that demonstrations must be near-optimal for effective training. Instead, the company emphasizes that data quality is best achieved at the distribution level, rather than focusing solely on individual trajectories. By collecting a diverse range of trajectories from a broad crowd, even if they are noisy and suboptimal, Axis believes that a robust policy can be developed during training.
The results from LIBERO-Plus showcase the effectiveness of continual pretraining on V1, leading to a substantial increase in success rates compared to a baseline model. As Axis gears up for V2, which will feature 1.2 million trajectories across 1,200 tasks, the team is confident in the scalability and generalization capabilities of their approach.
The dataset is just one component of Axis’s comprehensive data engine, which operates across four key data lines: Simulation, Egocentric, Loco-manipulation, and Human-gated DAgger post-training. Each line is designed to collect specific data tailored to different aspects of physical AI, with a focus on continuous improvement and learning.
Moving beyond open-sourcing simulation data, Axis collaborates with robot embodiment companies to develop customized data pipelines and model priors. By working closely with partners like Booster Robotics, Axis has demonstrated the effectiveness of their approach in real-world scenarios, achieving significant success rates with minimal real-robot demonstrations.
Axis Robotics is redefining the foundation of Physical AI by emphasizing the importance of an evolving data engine that adapts to the model’s needs. The company’s innovative approach, backed by a diverse team of researchers and experienced founders, is poised to make a significant impact in the field of robotics and AI.
For more information on Axis Robotics and their groundbreaking work, you can visit their Paper Link, Project Page, Dataset Link, and Github Codebase provided in the article.