AXIS: A Growable Community-Driven Data Engine for Scalable Robot Manipulation

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 $π_{0.5}$ 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.

• The paper presents a novel community-driven platform that transforms traditional robot manipulation datasets into an adaptable, continuously growing benchmark.
• It leverages browser-based teleoperation with automated task generation, rigorous data cleaning, and IsaacSim-driven augmentation to deliver high-fidelity trajectories.
• Empirical results highlight improved success rates and robustness, with performance scaling monotonically across various simulated perturbation axes.

AXIS: A Community-Driven Scalable Data Engine for Robot Manipulation

Motivation and Context

Effective robot manipulation policy learning critically depends on the availability of large-scale, diverse, and high-quality demonstration data. Historically, robot manipulation datasets have been hampered by expensive data collection procedures, reliance on specialized hardware, and rigid, centralized collection paradigms that tend to produce static benchmarks. This severely limits the ability to expand dataset coverage—across new tasks, environments, and contributors—in tandem with advances in model scale and capability.

AXIS ("A Growable Community-Driven Data Engine for Scalable Robot Manipulation" (2607.21588)) addresses these limitations by introducing an extensible platform for continuous, community-driven collection, refinement, and augmentation of robot manipulation data using browser-based teleoperation atop simulated environments. The central goal is to transform the data pipeline from a one-time, expert-controlled process into a continuously adaptive system that can scale with model innovations and the demands of general, robust robot manipulation.

System Architecture and Data Pipeline

AXIS is organized into three logically interconnected layers: infrastructure, dataset, and model.

Infrastructure Layer: This layer includes the browser-based MuJoCo-WASM teleoperation interface for broad, hardware-neutral access, enabling contributors worldwide to collect demonstrations with commodity input devices. Automated task generation parses high-level language instructions—backed by LLMs and asset normalization pipelines—to synthesize diverse, parameterized environment configurations and procedural success checkers for each new task. The system orchestrates community data collection via a web interface with instant feedback and consistent task deployment.

Dataset Layer: Raw community demonstrations are serialized in a unified trajectory schema capturing task metadata, robot and object states, actions, synchronized RGB-D observations, and success labels. The pipeline systematically validates, cleans, and refines trajectories. It employs automated consistency checks, task-specific physical validation, static-segment (hesitation) removal, temporal smoothing (Savitzky-Golay filtering), and temporal resampling to generate high-fidelity trajectories for learning. A subsequent IsaacSim-based augmentation step injects broad visual and physical diversity through randomization of lighting, camera pose, clutter layouts, dynamic parameters, and scene textures—enabling scalable robustness to environmental and sensor domain shift.

Model Layer: Policies—including both conventional visuomotor imitation learning and contemporary vision-language-action (VLA) models—are trained and evaluated on AXIS's data, using standardized task definitions and unified success criteria. The platform structures the dataset into "snapshots" (25%, 50%, 100%, etc.), enabling scalable benchmarking experiments where only the data volume varies under fixed protocol and architecture, facilitating controlled scaling studies.

Dataset Properties and Coverage

AXIS, as of the reported snapshot, contains:

• 207 distinct tabletop manipulation tasks with parameterized language instructions and environments.
• Over 50,000 human demonstration trajectories collected via web teleoperation.
• More than 60,000 augmented task or scene variants generated through scene and visual randomization.
• Rich annotations including synchronized third-person and end-effector (wrist) RGB streams, proprioceptive states, and detailed action traces.

Tasks span domains such as pick-and-place, stacking, sorting, articulated manipulation, tool use, pouring, insertion, and alignment, across seven distinct scene categories. Data growth is versioned, with each snapshot retaining compatibility for reproducibility and consistent longitudinal benchmarking.

Data Quality, Cleaning, and Augmentation

Data quality is maintained through:

• Multi-tier validation: Structural, numerical, physical plausibility, and replayable success according to ground-truth checkers.
• Trajectory refinement: Removal of idle or corrupted segments, temporal smoothing (yielding a 63.9% reduction in mean acceleration and 80.8% reduction in mean jerk relative to raw teleoperation), and upsampling to control-compatible frequencies.
• Augmentation: IsaacSim-driven stochastic scene rendering introduces variations in layout, clutter, lighting, camera pose, textures, and object parameters, producing broad distributions for policy generalization.

Crucially, augmentation directly targets axes of visual and geometric distributional shift encountered in robust real-world settings and systematic simulation benchmarks.

Policy Evaluation and Experimental Findings

AXIS serves as both a dataset and a benchmark for manipulation policy research in continual pretraining and scaling studies. The framework supports systematic evaluation under the LIBERO-Plus suite—which introduces structured perturbations including camera viewpoint, lighting, sensor noise, scene background, object layout, task language, and robot initial pose. The key results are:

• AXIS continual pretraining (on the full 100% snapshot) improves LIBERO-Plus success rates by 5.8% over the To.5 vanilla baseline and by 37.3% over a volume-matched RoboCasa365 continual pretraining control.
• Performance scales monotonically with AXIS data volume: Success rates progress from 84.7% (25%) to 85.7% (50%) to 88.8% (100%), with no clear saturation.
• Largest robustness gains observed for explicitly augmented axes:
  • Camera: +15.6%
  • Sensor noise: +16.6%
  • Layout: +3.1%
  • Robot pose: +5.1%
• AXIS outperforms equivalently-sized simulation datasets across nearly all LIBERO-Plus axes, highlighting the unique benefits of community-driven, diverse, and augmented data beyond mere scale.

Importantly, gains are not uniform but particularly marked for perturbation axes addressed by the augmentation pipeline, demonstrating direct transfer of simulated diversity to improved downstream policy robustness.

Broader Implications

Practical Significance: Community-driven, browser-based data collection lowers the barrier to contribution, enabling scale previously inaccessible without extensive physical robot fleets or institutional resources. Automated cleaning and augmentation make such data suitable for high-fidelity policy learning without manual curation overhead.

Theoretical and Research Impact: AXIS enables controlled scaling studies, disentangling the effects of data scale from model and protocol variables, essential for future research on dataset scaling laws and generalization in embodied agents. The platform supports rapid iteration and failure-driven expansion, aligning data evolution with model needs.

Transfer and Generalization: Robustness improvements realized through AXIS pretraining persist under held-out scenes, object layouts, and domain perturbations, evidencing direct generalization benefit from diversity and structured augmentation.

Future Directions: The present focus of AXIS is on simulated Franka-based tabletop manipulation. The paper acknowledges limitations in sim-to-real transfer and outlines concrete future expansions to multi-embodiment, richer sensor modalities, long-horizon and compositional tasks, active task-driven data growth, and real-world deployment. These would further solidify AXIS as an extensible, persistent data backbone for unified embodied AI research.

Conclusion

AXIS presents a scalable, community-first approach to robot manipulation data collection and benchmarking. Its architecture unifies automated task generation, web-based collection, rigorous cleaning, and IsaacSim-based augmentation in a pipeline designed for controlled scaling studies and continual growth. Empirical results support the conclusion that both the quantity and diversity of AXIS data, especially when coupled with purposeful augmentation, materially improve manipulation policy robustness under real-world-inspired perturbations. AXIS stands as a blueprint for the next generation of scalable, adaptive robotics datasets—critical infrastructure for advancing embodied generalist policies and lifelong robotic learning.