Infrastructure for Scalable Robot Learning

Arctivar Robotics Platform

A Cross-Simulator, Cross-Machine Infrastructure for Scalable Teleoperation and Generalizable Robot Learning

Arctivar Robotics Team
Arctivar Robotics Platform architecture

Arctivar Robotics Platform: A unified infrastructure bridging web-based teleoperation, GPU-accelerated realistic augmentation, and sim-to-real deployment.

Abstract

Robotic learning systems increasingly rely on large-scale demonstration data and realistic simulation for training and evaluation. However, existing workflows are often fragmented across simulators, operating systems, and compute environments: lightweight simulators enable broad access for teleoperation but lack high-fidelity rendering and physics, while GPU-based simulation stacks enable realism but are expensive and restrictive to access.

We present Arctivar Robotics Platform, a unified infrastructure that supports (i) multi-simulator development and execution, (ii) cross-machine orchestration spanning heterogeneous compute (web, Mac/Windows, and Linux GPU servers), and (iii) an end-to-end data pipeline from web-based teleoperation to photorealistic augmentation and downstream model training with sim-to-real deployment.

In Arctivar Robotics Platform, users teleoperate robots directly in a browser via a MuJoCo WebAssembly frontend, producing demonstrations without specialized hardware or compute. Demonstrations are then uploaded in a unified trajectory format and replayed on GPU-based Linux servers to generate realistic, domain-randomized rollouts using an IsaacSim-based augmentation backend. We further provide two production pipelines: (1) data cleaning and trajectory refinement, and (2) model training and sim-to-real evaluation, enabling rapid iteration from data to real-world deployment.

System Overview

System Overview

Platform Architecture

Explore the interconnected modules that make up the Arctivar Robotics Platform infrastructure, bridging lightweight data collection with heavy-duty simulation.

WASM UI
Task Generation
Resample Only
Smoothed & Resampled

Drag the slider to compare Resample Only vs. Smoothed & Resampled trajectory.

5x Speed

Trajectory Optimization Validation

Quantitative evaluation of trajectory optimization across LIBERO tasks. Our offline cleaning pipeline significantly reduces high-frequency artifacts (acceleration and jerk) from raw web-teleoperated data.

LIBERO TaskMean Acc. ↓Mean Jerk ↓Pos. Dev. (m)Removed Ratio
BeforeAfter (Ours)BeforeAfter (Ours)
Task 1: Place Black Bowl on Top of cabinet0.24580.07382.64251.33630.02356.46%
Task 2: Place Rear Butter in Cabinet Top Drawer and Close It0.30670.09433.73951.62080.06545.73%
Task 3: Place the black bowl on the plate0.35590.10115.13582.36340.02017.17%
Task 4: Place the black bowl on top of the cabinet0.32250.08594.67242.07850.01423.77%
Task 5: Place the frying pan on the stove0.18320.10552.47711.13740.01061.24%
Task 6: Place the moka pot on the stove0.18190.10902.52391.67840.00641.94%
Task 7: Turn on the stove0.26890.12953.82912.30170.01423.81%
Task 8: Close Cabinet Bottom Drawer0.15740.07411.91180.66590.00574.59%
Task 9: Place the black bowl into the cabinet’s bottom drawer0.15510.09131.86651.31930.01612.03%
Task 10: Place Wine Bottle on Wine Rack0.18410.11142.02981.00220.06254.86%
Task 11: Close Cabinet Top Drawer0.12240.06831.54240.58220.00532.62%
Task 12: Place the black bowl into the cabinet’s top drawer0.21090.14182.69972.16330.096111.4%
Task 13: Place Black Bowl on Plate0.16820.08162.22621.15520.25112.14%
Task 14: Place the black bowl on top of the cabinet0.13750.07361.84990.99240.01086.73%
Task 15: Place the right moka pot on the stove0.18900.11302.57681.49290.01343.19%
Task 16: Turn off the stove0.21490.12132.82681.58230.02198.93%
Average0.21280.09852.78441.46700.03984.79%

BibTeX

@misc{AxisWebInfra,
  author = {{AxisAIOrg}},
  title = {AxisWebInfra},
  year = {2026},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/AxisAIOrg/AxisWebInfra}},
}