Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Nimbus: A Unified Embodied Synthetic Data Generation Framework

Zeyu He, Yuchang Zhang, Yuanzhen Zhou, Miao Tao, Hengjie Li, Yang Tian, Jia Zeng, Tai Wang, Wenzhe Cai, Yilun Chen, Ning Gao, Jiangmiao Pang

TL;DR

Nimbus addresses fragmentation and instability in synthetic data pipelines for embodied AI by delivering a unified four layer architecture that decouples planning, rendering and storage. Through stage level decoupling, dynamic scheduling, global load balancing and backend renderer optimizations, Nimbus achieves end to end throughput improvements of two to three times and robust operation on large GPU clusters. The framework serves as the production backbone for the InternData suite and demonstrates a scalable, cross domain data synthesis capability. Future work includes expanding trajectory planners, automating scene generation, and enabling automated task configuration to further enhance downstream model generalization.

Abstract

Scaling data volume and diversity is critical for generalizing embodied intelligence. While synthetic data generation offers a scalable alternative to expensive physical data acquisition, existing pipelines remain fragmented and task-specific. This isolation leads to significant engineering inefficiency and system instability, failing to support the sustained, high-throughput data generation required for foundation model training. To address these challenges, we present Nimbus, a unified synthetic data generation framework designed to integrate heterogeneous navigation and manipulation pipelines. Nimbus introduces a modular four-layer architecture featuring a decoupled execution model that separates trajectory planning, rendering, and storage into asynchronous stages. By implementing dynamic pipeline scheduling, global load balancing, distributed fault tolerance, and backend-specific rendering optimizations, the system maximizes resource utilization across CPU, GPU, and I/O resources. Our evaluation demonstrates that Nimbus achieves a 2-3X improvement in end-to-end throughput compared to unoptimized baselines and ensuring robust, long-term operation in large-scale distributed environments. This framework serves as the production backbone for the InternData suite, enabling seamless cross-domain data synthesis.

Nimbus: A Unified Embodied Synthetic Data Generation Framework

TL;DR

Nimbus addresses fragmentation and instability in synthetic data pipelines for embodied AI by delivering a unified four layer architecture that decouples planning, rendering and storage. Through stage level decoupling, dynamic scheduling, global load balancing and backend renderer optimizations, Nimbus achieves end to end throughput improvements of two to three times and robust operation on large GPU clusters. The framework serves as the production backbone for the InternData suite and demonstrates a scalable, cross domain data synthesis capability. Future work includes expanding trajectory planners, automating scene generation, and enabling automated task configuration to further enhance downstream model generalization.

Abstract

Scaling data volume and diversity is critical for generalizing embodied intelligence. While synthetic data generation offers a scalable alternative to expensive physical data acquisition, existing pipelines remain fragmented and task-specific. This isolation leads to significant engineering inefficiency and system instability, failing to support the sustained, high-throughput data generation required for foundation model training. To address these challenges, we present Nimbus, a unified synthetic data generation framework designed to integrate heterogeneous navigation and manipulation pipelines. Nimbus introduces a modular four-layer architecture featuring a decoupled execution model that separates trajectory planning, rendering, and storage into asynchronous stages. By implementing dynamic pipeline scheduling, global load balancing, distributed fault tolerance, and backend-specific rendering optimizations, the system maximizes resource utilization across CPU, GPU, and I/O resources. Our evaluation demonstrates that Nimbus achieves a 2-3X improvement in end-to-end throughput compared to unoptimized baselines and ensuring robust, long-term operation in large-scale distributed environments. This framework serves as the production backbone for the InternData suite, enabling seamless cross-domain data synthesis.
Paper Structure (49 sections, 2 equations, 15 figures, 2 tables)

This paper contains 49 sections, 2 equations, 15 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. InternData-N1
  4. InternData-A1
  5. InternData-M1
  6. Related Work
  7. Navigation Data Generation
  8. Manipulation Data Generation
  9. Physical Teleoperation and Multi-Robot Aggregation.
  10. UMI-like Interfaces and Proxy Devices.
  11. Synthetic Data: Augmentation and Rule-Driven Generation.
  12. Summary: The Dataset-Centric Bottleneck.
  13. System Frameworks for Data Generation
  14. Architecture
  15. Stage Runner Layer
  16. ...and 34 more sections

Figures (15)

  • Figure 1: InternData-N1 pipeline
  • Figure 2: InternData-A1 pipeline
  • Figure 3: InternData-M1 pipeline
  • Figure 4: Nimbus Layered Architecture. The framework separates domain logic (Components Layer), workflow abstraction (Stage Runner Layer), control logic (Schedule Opt Layer), and execution runtime (Backend Opt Layer).
  • Figure 5: Example of Nimbus implementing data generation for different data pipelines via component combination
  • ...and 10 more figures