Model-base large scale parallel trajectory generation

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^*

^† Equal contribution ^* Corresponding author

Existing data generation methods suffer from exploration limits, embodiment gaps, and low signal-to-noise ratios, leading to performance degradation during self-iteration. We propose Seed2Scale, a self-evolving data engine that breaks 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 VLA model SuperTiny for robust parallel exploration, while a frozen Qwen3-VL-32B verifier autonomously filters trajectories — achieving a 209.15% performance improvement with no additional human annotations.

Paper PDF Demo Video RealMirror Platform BibTeX

Avg. Improvement

Seed Demos Required

Final Success Rate

26.3 Hz

SuperTiny Control Freq.

SCROLL

01 — Abstract

Breaking the
Data Bottleneck

A self-evolving framework that generates massive, high-quality robot trajectories from minimal human input, via heterogeneous model synergy.

The emergence of Vision-Language-Action (VLA) models is accelerating the evolution of Embodied AI. However, the performance of these models remains strictly contingent upon large-scale, high-quality expert demonstration trajectories, creating a critical "data scarcity" bottleneck that fundamentally obstructs the scaling of Generalist Embodied AI.

To address this, we propose Seed2Scale, a self-evolving data engine based on three synergistic roles: a lightweight SuperTiny collector (πsmall) for high-throughput parallel exploration, a frozen Qwen3-VL verifier (ΦVLV) for automated quality scoring, and a target SmolVLA model (πtarget) trained exclusively on curated high-quality trajectories.

By integrating these heterogeneous components, the engine transforms a sparse set of 4 human demonstrations into a continuous flow of verified training data. Experiments demonstrate a 209.15% average performance improvement, with the Can Stacking task alone achieving +778.67% relative improvement. Code, data, and model checkpoints will be publicly available.

02 — Overview Video

Display Video

A comprehensive demonstration of the Seed2Scale pipeline — from seed data collection to large-scale self-evolution and multi-task policy evaluation.

03 — Framework

The Seed2Scale Pipeline

From 4 boundary-point demonstrations to massive high-quality training data — a complete self-evolving ecosystem powering the target VLA model to new performance heights.

Seed2Scale self-evolving data engine pipeline overview

Fig. 1 — The Seed2Scale self-evolving data engine. Starting with as few as four human demonstrations, SuperTiny performs large-scale parallel rollouts. Qwen3-VL acts as a Verifier, scoring and filtering raw trajectories. The curated D_silver dataset trains SmolVLA through deep iterative learning — achieving robust scaling without massive manual annotations.

04 — Methodology

Heterogeneous Model Synergy

Three specialized roles working in concert, each exploiting distinct scale asymmetries to achieve what no single model can accomplish alone.

SuperTiny Collector

A lightweight 48M-parameter VLA model with ResNet-18 vision, T5-Small language encoder, and a Transformer decoder with temporal ensembling. Achieves 26.3 Hz inference — 3.6× faster than Diffusion Policy. Bootstraps robustly from just 4 seed trajectories without overfitting.

π_small · 48M params · 26.3 Hz

VLM-as-a-Verifier

Frozen Qwen3-VL-32B serves as the Vision-Language Verifier (VLV). For each trajectory τ, it receives task instruction, rollout video, and a reference demo, outputting a quality score S_τ ∈ [0, 10]. Only trajectories with S_τ ≥ γ enter D_silver, preventing data-poisoning spirals.

Φ_VLV · Qwen3-VL-32B · Frozen

SmolVLA Target Model

The final policy (π_target) trained via Conditional Flow Matching on curated D_silver. Its Action Expert architecture interleaves Cross-Attention (CA) over vision-language context with Self-Attention (SA) for temporal coherence — enabling stable closed-loop deployment on Agibot A2 and GR-1 robots.

π_target · 450M params · Flow Matching

Self-Evolution Loop

4 Seed Demos

Boundary-point teleoperation

SuperTiny

Parallel rollout N envs

VLV Filter

Score S ≥ γ threshold

D_silver

Curated trajectories

SmolVLA

Target model training

Next Iteration

D⁽ⁱ⁺¹⁾ = D⁽ⁱ⁾ ∪ D_silver

05 — Qualitative Results

Task Demonstrations

Direct comparison between the 4-point-only baseline and Seed2Scale-augmented policy across representative manipulation tasks.

Kitchen Cleanup · Task 1 BASELINE SEED2SCALE

4-Point Baseline

Seed2Scale (Ours)

Can Stacking · Task 4 BASELINE SEED2SCALE

4-Point Baseline

Seed2Scale (Ours)

Seed2Scale vs. MimicGen on GR-1 Tasks

Smoother Motion, Higher Success

Seed2Scale produces physically plausible, smooth trajectories. MimicGen's IK-based approach introduces severe high-frequency instability, yielding kinematically infeasible motions..

Wheel ManipulationMimicGen

Policy: 34.75% · Replay: 48.50%

Wheel ManipulationSeed2Scale

Policy: 93.25% ↑+168% · Replay: 67.86%

Cylinder GraspingMimicGen

Policy: 37.25% · Replay: 21.00%

Cylinder GraspingSeed2Scale

Policy: 66.00% ↑+77% · Replay: 86.96% ↑4×

06 — Quantitative Results

Performance Metrics

Comprehensive evaluation across multiple tasks demonstrating consistent, large-scale improvements with only 4 seed demonstrations.

+0%

Average relative improvement across all tasks

22.18% → 68.57%

+0%

Best single-task improvement (Can Stacking)

7.50% → 65.90%

Replay success over MimicGen (Cylinder Grasping)

21.00% → 86.96%

SuperTiny speed vs. Diffusion Policy

38ms vs 136ms inference

Agibot A2 Task Performance

Multi-task joint training · 4 seed demos per task · H20 training / RTX 5090 inference

Task	Seed Only	Seed2Scale (Ours)	Improvement
Kitchen Cleanup	24.63%	71.43%	+190.01%
Cup-to-Cup Transfer	23.50%	64.14%	+172.94%
Can Stacking	7.50%	65.90%	+778.67%
Air Fryer Manipulation	33.08%	72.82%	+120.13%
Average	22.18%	68.57%	+209.15%

GR-1 Tasks: Seed2Scale vs. MimicGen

Metric	Task	MimicGen	Seed2Scale	Δ
Policy Success	Cylinder Grasping	37.25%	66.00%	+77.18%
	Wheel Manipulation	34.75%	93.25%	+168.35%
	Average	36.00%	79.63%	+121.19%
Replay Success	Cylinder Grasping	21.00%	86.96%	+4.1×
	Wheel Manipulation	48.50%	67.86%	+39.92%
	Average	34.75%	77.41%	+122.77%

Trajectory Quality (lower = better)

Metric	Expert Demo	MimicGen	Seed2Scale
Total Variation ↓	1.32	3.68	1.34
Mean Abs. Jerk ↓	0.0063	0.0261	0.0047
HF Power Ratio (%) ↓	0.22	2.07	0.30

Motor joint position curves comparing Expert Demonstration, MimicGen and Seed2Scale across 4 joints

Fig. — Joint Motor Curve Comparison. Position trajectories (rad) for four representative joints over a full manipulation episode. Seed2Scale (orange dashed) closely tracks the Expert Demonstration (green dotted) in both amplitude and phase, while MimicGen (blue solid) diverges significantly — reflecting its higher Total Variation and Jerk scores in the table.

Ablation: Success Rate of SuperTiny vs SuperTiny-minus, Diffusion Policy and ACT across 7 self-evolution iterations

Fig. — Collector Ablation: Success Rate over Self-Evolution Iterations. SuperTiny (full model) consistently outperforms its ablated variant SuperTiny⁻ (without key components), Diffusion Policy, and ACT across all 7 iterations — plateauing near 50% while alternatives remain below 40% — confirming the critical role of SuperTiny's design choices in driving self-evolution.

Scaling: Target Model vs. Self-Evolution Iteration

Can Stacking — SmolVLA success rate across 8 iterations

Success Rate — Seed Only vs. Seed2Scale

Per-task comparison on Agibot A2

Kitchen Cleanup

Cup-to-Cup Transfer

Can Stacking

Air Fryer Manipulation

Seed Only Seed2Scale

07 — Architecture

Model Specifications

Scale asymmetry by design — a tiny collector enables massive parallel exploration, a large frozen VLM ensures quality, and a mid-size target model generalizes across tasks.

SuperTiny VLA

Small-Scale Collector · π_small

48M

Vision EncoderResNet-18 (3 cams)
Language EncoderT5-Small (frozen)
DecoderTransformer (cross-attn)
Action Chunk K30
Inference Time38.08 ms
Control Frequency26.3 Hz
EnsemblingExp. temporal (λ)
HardwareRTX 5090

Vision-Language Verifier

Large-Scale Evaluator · Φ_VLV

32B

Base ModelQwen3-VL-32B
ModeFrozen (no fine-tuning)
Input 1Task instruction g
Input 2Rollout video V_τ
Input 3Reference video V_ref
Output ScoreS_τ ∈ [0, 10]
RoleSuccess/Fail + Quality
Threshold γConfigurable per task

SmolVLA Target Model

Final Policy · Conditional Flow Matching · π_target

450M

Parameters450M
ObjectiveConditional Flow Matching
OptimizerAdamW
Learning Rate3 × 10⁻⁴

Training Steps100,000
Action Arch.CA + SA Action Expert
Training DataD_silver (curated)
DeploymentAgibot A2 / GR-1

SuperTiny VLA Architecture: Vision-State-Language Encoder, Multimodal Decoder, and Action Head with Temporal Ensembling

Fig. — SuperTiny VLA Architecture. Three camera inputs are encoded by a shared ResNet backbone into 192 visual tokens; robot joint states pass through a State MLP into state tokens; and task instructions are embedded via a frozen T5 encoder into language tokens. All three token streams are fused in a Multimodal Decoder (N layers of Self-Attention → Cross-Attention → FFN), producing a 30×128 action sequence refined by an Action Head with Temporal Ensembling (λ) for smooth 26.3 Hz closed-loop control.

08 — Multi-Task Demo

Four-Task Generalization

Multi-task joint training with cross-task knowledge transfer consistently achieves higher success rates than single-task training. The target model generalizes across diverse manipulation scenarios on the Agibot A2 robot platform.

Kitchen Cleanup

Pick up chip from the table, then place it into the basket. Tests precise grasping and placement from arbitrary positions beyond the 4 seed locations.

Success Rate: 71.43%

Cup-to-Cup Transfer

Pour berries from the right cup into the left cup. Requires fine motor control and understanding of granular transfer dynamics with non-uniform objects.

Success Rate: 64.14%

Can Stacking

Stack cans from both sides into the center with stable placement. Most dramatic improvement (+778%), requiring complex spatial reasoning far beyond the 4 seed demos.

Success Rate: 65.90%

Air Fryer Manipulation

Open the air fryer drawer and place food items inside. Tests multi-step manipulation with articulated objects, requiring precise door-handle grasping and coordinated placement.

Success Rate: 72.82%

10 — Acknowledgements

Special Thanks

AGIBot A2 for seed2scale.

AGiBot provides A2 robot model and open-source support

We extend our sincere gratitude to AGiBot for providing the Agibot A2 robot model and open-source support that made the large-scale experimental validation of Seed2Scale possible.

11 — Citation

BibTeX

If you find Seed2Scale useful in your research, please consider citing our work.

seed2scale.bib

@article{tai2026seed2scale,
  title   = {Seed2Scale: A Self-Evolving Data Engine for Embodied AI via Small to Large Model Synergy and Multimodal Evaluation},
  author  = {Tai, Cong and Zheng, Zhaoyu and Long, Haixu and Wu, Hansheng and Long, Zhengbin and Xiang, Haodong and Shi, Rong and Cui, Zhuo and Zhang, Shizhuang and Qiu, Gang and Wang, He and Li, Ruifeng and Liu, Biao and Sun, Zhenzhe and Shen, Tao},
  journal = {arXiv preprint arXiv:2603.08260},
  year    = {2026}
}

Seed2Scale

Breaking theData Bottleneck

Display Video

The Seed2Scale Pipeline

Heterogeneous Model Synergy

SuperTiny Collector

VLM-as-a-Verifier

SmolVLA Target Model

4 Seed Demos

SuperTiny

VLV Filter

Dsilver

SmolVLA

Next Iteration

Task Demonstrations

Smoother Motion, Higher Success

Performance Metrics

Agibot A2 Task Performance

GR-1 Tasks: Seed2Scale vs. MimicGen

Trajectory Quality (lower = better)

Model Specifications

SuperTiny VLA

Vision-Language Verifier

SmolVLA Target Model

Four-Task Generalization

Kitchen Cleanup

Cup-to-Cup Transfer

Can Stacking

Air Fryer Manipulation

Full Paper

Special Thanks

AGiBot provides A2 robot model and open-source support

BibTeX

Breaking the
Data Bottleneck

D_silver