Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Seed2Scale: A Self-Evolving Data Engine for Embodied AI via Small to Large Model Synergy and Multimodal Evaluation

Cong Tai, Zhaoyu Zheng, Haixu Long, Hansheng Wu, Zhengbin Long, Haodong Xiang, Rong Shi, Zhuo Cui, Shizhuang Zhang, Gang Qiu, He Wang, Ruifeng Li, Biao Liu, Zhenzhe Sun, Tao Shen

TL;DR

Seed2Scale is a self-evolving data engine that overcomes the data bottleneck through a heterogeneous synergy of small-model collection, large-model evaluation, and target-model learning, providing a scalable and cost-effective pathway for the large-scale development of Generalist Embodied AI.

Abstract

Existing data generation methods suffer from exploration limits, embodiment gaps, and low signal-to-noise ratios, leading to performance degradation during self-iteration. To address these challenges, we propose Seed2Scale, a self-evolving data engine that overcomes the data bottleneck through a heterogeneous synergy of "small-model collection, large-model evaluation, and target-model learning". Starting with as few as four seed demonstrations, the engine employs the lightweight Vision-Language-Action model, SuperTiny, as a dedicated collector, leveraging its strong inductive bias for robust exploration in parallel environments. Concurrently, a pre-trained Vision-Language Model is integrated as a Verifer to autonomously perform success/failure judgment and quality scoring for the massive generated trajectories. Seed2Scale effectively mitigates model collapse, ensuring the stability of the self-evolution process. Experimental results demonstrate that Seed2Scale exhibits signifcant scaling potential: as iterations progress, the success rate of the target model shows a robust upward trend, achieving a performance improvement of 131.2%. Furthermore, Seed2Scale signifcantly outperforms existing data augmentation methods, providing a scalable and cost-effective pathway for the large-scale development of Generalist Embodied AI. Project page: https://terminators2025.github.io/Seed2Scale.github.io

Seed2Scale: A Self-Evolving Data Engine for Embodied AI via Small to Large Model Synergy and Multimodal Evaluation

TL;DR

Seed2Scale is a self-evolving data engine that overcomes the data bottleneck through a heterogeneous synergy of small-model collection, large-model evaluation, and target-model learning, providing a scalable and cost-effective pathway for the large-scale development of Generalist Embodied AI.

Abstract

Existing data generation methods suffer from exploration limits, embodiment gaps, and low signal-to-noise ratios, leading to performance degradation during self-iteration. To address these challenges, we propose Seed2Scale, a self-evolving data engine that overcomes the data bottleneck through a heterogeneous synergy of "small-model collection, large-model evaluation, and target-model learning". Starting with as few as four seed demonstrations, the engine employs the lightweight Vision-Language-Action model, SuperTiny, as a dedicated collector, leveraging its strong inductive bias for robust exploration in parallel environments. Concurrently, a pre-trained Vision-Language Model is integrated as a Verifer to autonomously perform success/failure judgment and quality scoring for the massive generated trajectories. Seed2Scale effectively mitigates model collapse, ensuring the stability of the self-evolution process. Experimental results demonstrate that Seed2Scale exhibits signifcant scaling potential: as iterations progress, the success rate of the target model shows a robust upward trend, achieving a performance improvement of 131.2%. Furthermore, Seed2Scale signifcantly outperforms existing data augmentation methods, providing a scalable and cost-effective pathway for the large-scale development of Generalist Embodied AI. Project page: https://terminators2025.github.io/Seed2Scale.github.io
Paper Structure (18 sections, 10 equations, 6 figures, 4 tables)

This paper contains 18 sections, 10 equations, 6 figures, 4 tables.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Vision-Language-Action
  4. Automated Data Generation and Augmentation
  5. Vision-Language Models
  6. Seed2Scale
  7. SuperTiny: Small-Scale VLA Collector
  8. VLM-as-a-Verifier
  9. Seed-to-Scale Bootstrapping
  10. Target Model Training
  11. Experiments
  12. Experimental Setup
  13. Can Seed2Scale significantly improve VLA policy performance with minimal seed data
  14. Does the self-evolution mechanism exhibit consistent scaling performance across iterations
  15. How does Seed2Scale compare to existing data augmentation methods
  16. ...and 3 more sections

Figures (6)

  • Figure 1: Architecture of the SuperTiny VLA. The model integrates vision (ResNet), language (T5), and robot state encodings into a unified conditioning memory $\mathcal{M}_t$, which is processed by a lightweight Transformer decoder via cross-attention to predict action chunks. Temporal ensembling (Eq. 4) ensures smooth and consistent control.
  • Figure 2: (a) GR-1 robot tasks used for comparison with MimicGen and Seed2Scale. (b) Agibot A2 robot tasks for Seed2Scale evaluation.
  • Figure 3: Scaling performance of the target model (SmolVLA) across self-evolution iterations.
  • Figure 4: Comparison of robot action curves generated by expert demonstration, MimicGen, and Seed2Scale.
  • Figure 5: Performance comparison of SuperTiny, ACT, and Diffusion Policy as data collectors across self-evolution iterations. SuperTiny$^-$ denotes the variant without VLV-based quality filtering.
  • ...and 1 more figures