Capture • Dataset • Evaluation

AXIS

A Growable Community-Driven Data Engine
for Scalable Robot Manipulation

Mengfei Zhao*,‡, Dihong Huang1,2*, Yikai Tang2*, Peihao Li2*, Mingxuan Yan3*, Ruiqi Zhuang1*, Yanjia Huang4*, Jie Wang1,5, Hai Zhai1, Tony Zhou1,6, Rui Zhang1,7, Zhexi Luo1,8, Yuchen Huang1,7, Jianfei Yang9‡, Jiachen Li3‡

1Axis Robotics 2University of California, Berkeley 3Georgia Institute of Technology 4Texas A&M University 5Johns Hopkins University 6University of Pennsylvania 7University of Michigan 8National University of Singapore 9Nanyang Technological University

*Equal contribution    ‡zhaomengfei248@gmail.com, jianfei.yang@ntu.edu.sg, jiachen_li@gatech.edu

[paper] [Dataset (Hugging Face)] [Platform] Dataset Browser

Real-time statistics

1,530,275
Trajectories
1,816
Tasks
13,645 h 7 m
Trajectory duration

Data is verified every hour. Task growth uses the latest complete UTC day.

AXIS system overview showing task generation, teleoperation, data cleaning, augmentation, policy learning, and real-world validation.

AXIS system overview: task generation, browser-based teleoperation, offline data processing, simulation augmentation, policy learning, and real-world validation.

Abstract

Learning effective robot manipulation policies requires diverse, high-quality demonstrations, yet existing data pipelines are often difficult to scale because they rely on specialized hardware, centralized operators, or fixed task suites. We present AXIS, a growable community-driven data engine and benchmark for scalable robot learning, which enables browser-based teleoperation for large-scale demonstration collection, automatically generates and validates new manipulation tasks, and transforms community-collected demonstrations into training-ready data through automated success checking, quality filtering, trajectory smoothing, and visual and physics-based augmentation.

The AXIS dataset currently contains 207 diverse tasks and 50K+ trajectories. Meanwhile, AXIS organizes data into task snapshots and evaluates policies with a systematic held-out protocol. We compare vision-language-action (VLA) policies under a unified AXIS evaluation suite and analyze scaling behavior across different data volumes.

Continual pretraining on AXIS substantially improves the overall success rate of π0.5 by 5.8%, outperforms the model pretrained on RoboCasa365 by 37.3%, and exhibits consistent scaling with increasing data volume, with the largest gains observed under layout, sensor-noise, and camera perturbations.

Scalable Robot Data Collection

Collection stays lightweight in the browser; validation, rendering, augmentation, and learning run on the backend.

The AXIS collection layer separates latency-sensitive teleoperation from compute-heavy data processing. Contributors operate a Franka Research 3 arm in a MuJoCo-WASM browser frontend with commodity input devices, while uploaded trajectories are routed to backend machines for validation, cleaning, replay, rendering, augmentation, and export.

Data engine stages

  1. 01 Task generation and asset normalization
  2. 02 Web-based MuJoCo-WASM teleoperation
  3. 03 Success checking and trajectory upload
  4. 04 Filtering, hesitation removal, smoothing, and resampling
  5. 05 IsaacSim visual and physics augmentation
  6. 06 VLA training, fixed-protocol evaluation, and real-world rollout

Browser teleoperation

Contributors collect demonstrations without local simulator installation, GPUs, or specialized robot hardware.

Unified trajectory format

Each episode stores task metadata, embodiment, simulator version, observations, states, actions, and success information.

Quality refinement

Backend validation removes corrupted records, failed tasks, idle segments, jitter, and nonphysical discontinuities.

Realistic augmentation

Scene, camera, lighting, texture, pose, friction, mass, and dynamics variations expand the training distribution.

Browser-based teleoperation sessions showing remote task execution through the AXIS collection interface.
The AXIS Franka Dataset

AXIS is a tabletop manipulation dataset built around a Franka Research 3 robot with a parallel-jaw gripper. Each task includes a language instruction, a parameterized simulation scene, task assets, and a structured success checker. Each trajectory provides synchronized robot states, object states, robot actions, task metadata, success labels, and multi-view RGB-D observations.

Dataset snapshot

Tasks
207
Trajectories
50K+
Task/scene variants
60K+
Robot
Franka Research 3
Cameras
Third-view + wrist RGB-D
Evaluation
Fixed held-out protocol
AXIS dataset overview showing dataset scale, trajectory schema, augmentation, data growth, and task distribution.
Dataset overview covering scale, trajectory schema, validation and refinement, augmentation, growth snapshots, and scene/skill distribution.
Browser collection workflow for a manipulation task, including on-screen controls and simulator feedback.
Trajectory refinement

Raw community demonstrations may contain hesitation, jitter, low-frequency sampling artifacts, or invalid transitions. AXIS filters and refines these demonstrations before converting them into policy-learning data.

Effect of data refinement on teleoperation trajectory quality.
Data Version Sampling Rate Mean Acceleration Mean Jerk Replay Success
Raw Teleoperation 5.0 Hz 1.3539 11.5899 100.0%
Smoothed 5.0 Hz 0.6382 2.9160 91.4%
Smoothed + Resampled 20 Hz 0.4885 2.2243 86.2%
Augmentation diversity across example AXIS manipulation tasks.
Augmentation diversity across example tasks, covering randomized appearance, lighting, environments, and object poses.
A living benchmark

Instead of treating the dataset as a one-time release, AXIS organizes training data into progressively larger snapshots. AXIS-25%, AXIS-50%, AXIS-100%, and future versions preserve shared data format, task definitions, success checkers, rollout budgets, and evaluation tasks, so scaling studies can isolate the effect of training coverage.

AXIS-25%low-volume scaling point
AXIS-50%mid-scale task coverage
AXIS-100%current full snapshot
AXIS-...future growable releases
Grid of augmented tabletop simulation scenes with varied backgrounds and object layouts.
Additional augmented tabletop scenes with varied backgrounds, lighting, surface textures, object layouts, and camera viewpoints.
Simulation scenes spanning kitchen, tabletop, mug, and object manipulation tasks.
Scene and object diversity across simulated Franka manipulation tasks.

What the dataset contains

language instruction task assets success checker robot states object states robot actions third-view RGB-D wrist RGB-D metadata failure labels
Simulation demos
Same task with different augmentation settings.
Different task setup with varied scene assets and initial conditions.
Experiments

The experiments evaluate whether AXIS pretraining improves downstream LIBERO-Plus robustness for π0.5, whether the improvement scales with AXIS data volume, and which perturbation axes benefit most from AXIS-style augmentation and task diversity.

Headline result. π0.5 + AXIS-100% reaches 88.8 overall LIBERO-Plus success, compared with 83.9 for vanilla π0.5 and 57.5 for a RoboCasa-matched simulation baseline.
Main result on LIBERO-Plus.
Pretraining # demos Overall ↑ Cam. Light Noise B.G. Layout Lang. Robot
π0.5 vanilla 0 83.9 72.5 98.2 82.5 94.4 82.9 89.6 74.4
π0.5 + AXIS-25% 0.25 NAXIS 84.7 77.5 91.2 86.2 100.0 85.5 81.8 76.9
π0.5 + AXIS-50% 0.50 NAXIS 85.7 68.8 98.2 91.2 96.3 88.2 84.4 79.5
π0.5 + AXIS-100% (ours) NAXIS 88.8 83.8 96.5 96.2 98.1 85.5 88.3 78.2
π0.5 + RoboCasa-matched NAXIS 57.5 35.2 79.5 63.2 81.7 68.0 49.2 39.4
Per-perturbation robustness

Sensor Noise

+13.7

Largest gain over vanilla; robustness improves under photometric observation shifts.

Camera

+11.3

Viewpoint variation transfers to LIBERO-Plus camera shifts.

Robot Pose

+3.8

AXIS helps beyond explicitly randomized perturbation axes.

Background

+3.7

Scene and texture randomization improves background robustness.

AXIS-100% improvements by perturbation axis.
Axis π0.5 vanilla + AXIS-25% + AXIS-100% + RoboCasa-m. ∆van AXIS−RC
Camera 72.5 77.5 83.8 35.2 +11.3 +48.6
Light 98.2 91.2 96.5 79.5 −1.7 +17.0
Sensor Noise 82.5 86.2 96.2 63.2 +13.7 +33.0
Background 94.4 100.0 98.1 81.7 +3.7 +16.4
Layout 82.9 85.5 85.5 68.0 +2.6 +17.5
Language 89.6 81.8 88.3 49.2 −1.3 +39.1
Robot 74.4 76.9 78.2 39.4 +3.8 +38.8
Overall 83.9 84.7 88.8 57.5 +4.9 +31.3
Real-world rollouts
Third-view rollout with tabletop grocery objects. (x10 speed)
Third-view rollout under the same real-world setup. (x10 speed)
Wrist-view rollout with close-up observations. (x10 speed)
Wrist-view rollout with top-down observations. (x10 speed)
Real-world Franka rollout videos under third-view and wrist-view observations.
BibTeX
BibTeX 
@article{zhao2026axis,
  title   = {AXIS: A Growable Community-Driven Data Engine for Scalable Robot Manipulation},
  author  = {Mengfei Zhao and Dihong Huang and Yikai Tang and Peihao Li and Mingxuan Yan and Ruiqi Zhuang and Yanjia Huang and Jie Wang and Hai Zhai and Tony Zhou and Rui Zhang and Zhexi Luo and Yuchen Huang and Jianfei Yang and Jiachen Li},
  journal = {arXiv preprint arXiv:2607.21588},
  year    = {2026},
  url     = {https://arxiv.org/abs/2607.21588}
}