- Published on
A Browser-Based Data Engine With 207 Robot Manipulation Tasks and 50,129 Trajectories
- Authors
- Name
- DeepHorus
Axis Robotics has launched AXIS, a browser-based data engine designed to accelerate robot manipulation research by transforming how demonstration datasets are collected. This novel platform moves the process out of centralized labs and into any web browser, aiming to address the persistent bottleneck where dataset growth lags far behind the advancement of AI models.
Historically, robot manipulation datasets have been collected by expert operators in specialized labs, processed offline, and released as fixed benchmarks. This closed and centralized approach limits scalability and continuous improvement. The AXIS system presents a different model, treating the dataset as a living resource that continuously expands through community contribution.
The core innovation is a split architecture. Contributors use a browser interface to teleoperate a simulated Franka Research 3 robot within a MuJoCo WebAssembly frontend. The computationally intensive tasks of physics rendering, training, and evaluation are handled by a powerful backend powered by eight NVIDIA RTX 4090 and eight A100 GPUs. This design ensures high-quality simulation while keeping access open and lightweight for users.
The current released snapshot on Hugging Face is a massive repository containing 207 tasks, 50,129 trajectories, and over 60,000 task or scene variants. The dataset is gated for non-commercial academic use and totals 2.36 TB. Each trajectory includes rich metadata, simulator states, actions, success labels, and RGB-D observations, with contributions credited to over 70,000 community members.
A rigorous cleaning and validation pipeline ensures data quality. Samples are filtered for static motion, smoothed, and resampled to a higher frequency. Cleaned trajectories are then replayed in NVIDIA Isaac Sim to generate photorealistic observations with randomized scenes and camera views. The paper honestly reports trade-offs, such as a reduction in mean acceleration but an overall high replay success rate of 86.2%.
Preliminary results demonstrate the system's potential. When used for pretraining, the AXIS dataset helped a Ο0.5 model achieve 88.8% accuracy on the LIBERO-Plus benchmark, significantly outperforming a baseline model's 83.9%. The greatest improvements were seen in tasks involving sensor noise, where the augmentation pipeline was most active. Scaling from 25% to 100% of the data consistently improved performance.
The platform's training code is publicly available as a patch layer, and the teleoperation platform is live. However, no pre-trained policy checkpoints have been released. This release provides a powerful new tool and open-ended dataset for the robotics research community to build upon, signaling a shift toward more collaborative and scalable data collection methods for training advanced robot policies.
