Seed2Scale: A Self-Evolving Data Engine with Parallel Worlds Expansion for Scalable Robot Learning

The gist: Seed2Scale is a self-evolving data engine that breaks the data bottleneck in Embodied AI through a heterogeneous synergy of “small-model collection, large-model evaluation, and target-model learning”.

How it works

Seed2Scale addresses the limitations of existing data generation methods by employing a heterogeneous synergy architecture consisting of “small-model collection, large-model evaluation, and target-model learning” (Page 1). This framework is designed to overcome the data bottleneck in Embodied AI through a self-evolving process (Page 2). The core idea is to decouple the processes of exploration, verification, and final policy learning across models of disparate scales (Page 3).

The synergy is established through three specialized roles:

1. Small-Scale Collector (πsmall): A lightweight VLA model, SuperTiny, acts as a dedicated data collector (Page 3). This model leverages its strong inductive bias for robust exploration in parallel environments to bootstrap effectively from minimal seed data without the overfitting risks typical of larger architectures (Page 2).

2. Large-Scale Verifier (ΦV LV): A frozen, pre-trained Vision-Language Model (Qwen3-VL [16], [17]) serves as a Vision-Language Verifier (VLV) to autonomously score and filter generated trajectories by performing success/failure judgment and quality scoring (Page 3). This mechanism is crucial for averting the vicious cycle of model collapse caused by failed and low-quality data, ensuring stable and robust self-evolution (Page 2).

3. Target Model (πtarget): The final model, SmolVLA [2], is trained on the curated high-quality dataset Dsilver (Page 3). This decoupled training ensures that the target model focuses on distilling verified motion priors and semantic-action correlations, driving a continuous increase in its capabilities from basic actions to complex skills (Page 3).

Data Collection and Curation Pipeline

The engine initiates from as few as four seed demonstrations, which correspond to the four corner positions of the tabletop workspace (Page 3). The self-evolving process follows a recursive data expansion loop:

1. At iteration i, the lightweight collector π(i) small is trained on dataset D(i) (initialized as D(0) = Dseed), then deployed in Nenv parallel environments to generate raw trajectories T(i) raw (Page 3).

2. The VLV scores each trajectory τj (Sec. III-B), and only those exceeding a quality threshold γ are retained in the curated dataset Dsilver (Page 3).

3. The dataset is then augmented for the next iteration: D(i+1) = D(i) ∪ D(i) silver (Page 3).

Model Architectures and Training Objectives

The SuperTiny VLA model is designed for efficiency, utilizing a heterogeneous encoding strategy where Visual features are extracted via a ResNet-18 backbone and projected through a 1×1 convolution (Page 4). It employs exponential temporal ensembling over overlapping action chunks to ensure smooth control during high-frequency rollouts (Page 4). The target model, SmolVLA, is trained via Conditional Flow Matching, which learns a vector field vθ that maps noise into structured action sequences (Page 3). To ensure temporally coherent action chunks At = (at,..., at+K), SmolVLA employs an Action Expert architecture that interleaves Cross-Attention with Self-Attention to model temporal dependencies within the action sequence (Page 4).

Experimental Validation and Scaling Performance

The experimental results demonstrate significant scaling potential: as iterations progress, the success rate of the target model shows a robust upward trend, achieving a relative performance improvement of 209.15% (Page 2). For instance, on the Can Stacking task, the success rate increased from an initial 7.50% to 65.90% (Page 6). Seed2Scale significantly outperforms existing data augmentation methods like MimicGen, achieving a 4× improvement in replay success for Cylinder Grasping (Page 7). Furthermore, trajectory quality metrics show that Seed2Scale produces trajectories with human-like smoothness, significantly outperforming MimicGen in terms of Total Variation and Mean Absolute Jerk (Page 8).

Conclusion

Seed2Scale successfully transforms a sparse set of four human demonstrations into a continuous flow of high-quality, verified training data by orchestrating the synergy between the lightweight SuperTiny collector and the VLV (Page 8). This approach effectively mitigates model collapse and enables Generalist Embodied AI by overcoming the traditional data bottleneck (Page 2). The framework achieves a substantial performance leap, validating its viability for scalable robot learning (Page 8).

1. Bold Header: Cost-Efficient Data Generation

Seed2Scale enables largescale data generation from as few as four initial human demonstrations, significantly mitigating the reliance on manual data acquisition in Embodied AI. This allows for a continuous flow of high-quality, verified training data from minimal seed input rather than requiring massive manual annotations or expensive expert trajectories.

2. Bold Header: Robust Self-Evolution

The system employs a recursive loop where the lightweight collector π(i) small is trained on dataset D(i) (initialized as D(0) = Dseed), then deployed in Nenv parallel environments to generate raw trajectories. This process ensures that the target model learns from an ever-expanding, high-quality dataset, leading to a robust upward trend in performance across self-evolution iterations.

3. Bold Header: VLM-Guided Quality Control

The integration of the Vision-Language Model as a verifier addresses data noise by allowing it to autonomously perform success/failure judgment and quality scoring for the massive generated trajectories. This mechanism ensures that only trajectories exceeding a quality threshold γ are admitted into Dsilver, effectively averting the vicious cycle of model collapse caused by lowquality data contamination.

4. Bold Header: High-Fidelity Trajectory Synthesis

The combination of SuperTiny's parallelized rollouts and exponential temporal ensembling enables the collector to generate thousands of diverse trajectories without manual intervention while ensuring smooth control via weighted averaging. This results in trajectories that exhibit human-like smoothness, significantly outperforming MimicGen, evidenced by a lower Total Variation and Mean Absolute Jerk compared to existing methods.

5. Bold Header: Superior Policy Learning

The target model, SmolVLA, is trained using Conditional Flow Matching rather than standard behavior cloning, which learn[s] to denoise through iterative refinement. This training approach allows the model to learn from the curated dataset Dsilver and achieve a significant leap in performance (up to 209.15% improvement), enabling it to master complex skills like Can Stacking far exceeding the information density of initial demonstrations.

Existing data generation methods for robot learning suffer from limited exploration, embodiment gaps, low signal-to-noise ratios, and domain shifts, leading to performance degradation during self-iteration and poor generalization to unseen scenes. To address these challenges, we propose Seed2Scale, a self-evolving data engine with parallel worlds expansion. Starting with as few as four seed demonstrations, Seed2Scale first executes a self-evolution stage driven by a heterogeneous synergy of "small-model collection, large-model evaluation, and target-model learning". Specifically, the lightweight Vision-Language-Action (VLA) model, SuperTiny, serves as a dedicated data collector for robust exploration. Concurrently, a pretrained Vision-Language Model (VLM) functions as a verifier to autonomously score and filter trajectories, supporting stable self-evolution in the evaluated tasks without performance collapse. Furthermore, Seed2Scale introduces a parallel worlds stage, projecting self-evolved trajectories into different environments of the same task to generate more diverse data and enhance adaptability to unseen scenes, including real-world environments. Experimental results demonstrate that Seed2Scale exhibits significant scaling potential: as iterations progress, the success rate of the target model shows a consistent upward trend, significantly outperforming the seed baseline. Notably, Seed2Scale achieves a remarkable 75.38% success rate in zero-shot real-world evaluations, where baseline methods fail completely (0%). Project page: https://terminators2025.github.io/Seed2Scale.github.io