# Axis Robotics Releases Largest Open-Source Franka Arm Simulation Dataset Dataset with 50,000+ trajectories aims to redefine Physical AI training paradigms. - **Source:** [NewsBTC](https://brifing.si/source/1757e8d5-352f-4479-994f-79c6e6161b17) - **Published:** 2026-09-04 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/b38ec9ac-49c4-4a10-bfde-18ba9053848a ## Breaking Records in Robotics Data Axis Robotics' V1 dataset contains over 50,000 human-teleoperated trajectories across 207 manipulation tasks, making it the largest open-source Franka arm simulation collection. The dataset includes 60,000+ scene variants and has been downloaded 160,000+ times from Hugging Face, setting a new benchmark for accessible robotics research materials. Unlike curated 'clean data' approaches, Axis intentionally includes noisy, suboptimal trajectories from distributed contributors to prove their distribution-level quality thesis. ## Performance That Speaks Volumes Benchmarks show continual pretraining on V1 improved π0.5 performance from 83.9% to 88.8% success rates. The dataset outperformed volume-matched RoboCasa baselines by 37.3%, with consistent improvements as training data scaled from 25% to 100% utilization. Most significant gains appeared under camera and sensor perturbations - exactly the randomization axes Axis built into their data generation process. These results validate their approach of prioritizing diversity over curated perfection in training data. ## The Four-Pillar Data Engine Axis operates a compounding data generation system across four channels: simulation (4.7M+ trajectories), egocentric capture (200k+ hours from real environments), loco-manipulation (500+ humanoid hours), and human-gated DAgger correction. Unlike static datasets, their engine uses model failures to determine next collection priorities. All data is recorded on-chain via Base for provenance, with contributors rewarded for quality - creating a closed-loop system that turns failures into improvements. ## From Simulation to Real-World Impact With Booster Robotics, Axis demonstrated their approach's effectiveness by creating a digital twin workspace and collecting 42,000 simulation episodes. The resulting model prior achieved 87.5% success with just 30 real-robot demos, compared to 37.5% for standard π0.5. This 2.3x improvement with 50% fewer real-world demonstrations showcases how simulation data can accelerate real-world deployment timelines while reducing costs. ## Industry Adoption and Partnerships Axis works with embodiment companies (Feagine Robotics), model developers (Manycore Tech), and automotive giants (Lotus Cars, Geely Auto). They also power on-chain robotics networks like BitRobot (Solana) and OpenRoboto (Bittensor). Their $12M seed funding round was led by Hack VC with participation from Nomad Capital and Pi Network Ventures, signaling strong investor confidence in their vertically integrated Physical AI approach. ## Academic Foundations The project draws on research from UC Berkeley, Johns Hopkins, and University of Michigan, with advisory from Georgia Tech's Jiachen Li. The founding team includes researchers from top institutions (CMU, Georgia Tech, SJTU) alongside serial entrepreneurs who've scaled platforms to 30M+ users. This blend of academic rigor and practical scaling experience informs their unique approach to robotics data generation. ## What's Next: V2 and Beyond Axis is already developing V2, scaling to 1.2 million trajectories across 1,200 tasks. Early results show suboptimal simulation data can train robust policies across multiple VLA models. The team emphasizes cross-embodiment generalization as a key focus area, aiming to create transferable skills rather than task-specific solutions. Their daily data generation capacity now exceeds 4,000 hours of new egocentric capture. ## Redefining Physical AI's Future Axis challenges the robotics industry's 'clean data only' assumption, proving that diverse, noisy data at scale can outperform carefully curated datasets. Their compounding engine approach - where every training round informs the next collection priority - represents a paradigm shift from static datasets to dynamic, continuously improving systems. As founder Chris Feng states, this closed-loop process turns failures into progress, creating truly compounding AI advancement. ## Sources covering this story - https://www.newsbtc.com/news/axis-robotics-open-sources-one-of-the-largest-franka-arm-simulation-datasets-for-physical-ai/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._