Axis Sim Dataset V1
Axis Sim Dataset V1 is an open-source synthetic robot manipulation dataset maintained by Hong Kong based Axis Robotics, built for pretraining generalist physical AI policies. Spanning over 50,000 human-teleoperated simulation trajectories across 207 manipulation tasks and more than 60,000 scene variations, the corpus tests whether distributed, noisy simulation data can drive embodied representation learning. Trajectories feature a simulated Franka Research 3 robotic arm executing tasks such as pick-and-place, stacking, pouring, tool use, and articulated-object interactions via the web-based Axis Hub interface. Available openly on Hugging Face alongside training code on GitHub, the dataset achieved over 160,000 community downloads and demonstrated measurable empirical improvements when fine-tuning policies like π0.5 against RoboCasa baselines. Designed specifically for robotics researchers, AI engineers, and foundation model teams, Axis Sim Dataset V1 provides scalable, accessible synthetic demonstrations to advance robotic foundational manipulation models without demanding clean expert-only data.
| Year | 2026 |
|---|---|
| Trajectories | 50,000 |
| Embodiments | Franka Research 3 |
| Task categories | manipulation, pick-and-place |
| License | open-source |
| Access | open — commercial use permitted |
| Maintainer | Axis Robotics |
| Origin country | HK |
Overview
Axis Sim Dataset V1 is a large-scale open-source simulation manipulation dataset developed by Axis Robotics in collaboration with academic researchers from UC Berkeley, Johns Hopkins University, and the University of Michigan. The dataset was created to test whether noisy, crowd-sourced simulation demonstrations can successfully support embodied pretraining without relying on immaculate, expert-only trajectories.
Collection Methodology
Data collection was conducted through Axis Hub, a browser-based teleoperation platform that enables a globally distributed contributor crowd to manipulate simulated environments. Demonstrations were gathered across 207 distinct manipulation tasks involving over 60,000 procedural scene variations on a simulated Franka Research 3 arm. Task activities encompass pick-and-place operations, object stacking, liquid pouring, articulated-object manipulation, and basic tool use.
Benchmarks and Use Cases
Axis Sim Dataset V1 is engineered as an embodied pretraining base for robotics foundation models (such as π0.5). In LIBERO-Plus evaluation benchmarks, continual pretraining on V1 improved π0.5 success rates from 83.9% to 88.8%, surpassing volume-matched RoboCasa365 baselines by 37.3%. The full training codebase, model checkpoints, and datasets are publicly open on Hugging Face and GitHub.
Frequently asked questions
What robot embodiment is simulated in Axis Sim Dataset V1?
Axis Sim Dataset V1 focuses on a simulated Franka Research 3 robotic arm.
How many trajectories and tasks are included in Axis Sim Dataset V1?
The dataset contains more than 50,000 human-teleoperated simulation trajectories spanning 207 manipulation tasks and over 60,000 scene variants.
Where can the dataset and training code be accessed?
The dataset is hosted on Hugging Face (axisrobotics/Franka-Dataset), the training codebase is available on GitHub (AxisAIOrg/Axis-V1-Training), and documentation is hosted on the AXIS-V1 project page.
What is the primary scientific hypothesis behind Axis Sim Dataset V1?
The dataset tests the hypothesis that data quality resides at the distribution level rather than the individual trajectory, showing that uncorrelated noise from crowdsourced teleoperation averages out during large-scale pretraining.
Which foundation models have been evaluated using this dataset?
The dataset has been evaluated using continual pretraining on the π0.5 model, outperforming volume-matched baselines on the LIBERO-Plus benchmark.