Self-Evolving Data Engine · Parallel Worlds Expansion

Seed2Scale

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

Cong Tai^†·Zhaoyu Zheng^†·Haixu Long^†·Hansheng Wu^†·Jialiang Wang·Zhengbin Long
Haodong Xiang·Shuyuan Wang·Rong Shi·Zhuo Cui·Shizhuang Zhang
He Wang·Biao Liu·Zhenzhe Sun·Tao Shen^*

ZTE Corporation, China
^† Equal contribution ^* Corresponding author

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. 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", then projects self-evolved trajectories into parallel worlds for visual generalization — achieving a 75.38% success rate in zero-shot real-world evaluations, where baseline methods fail completely (0%).

Paper PDF Demo Video RealMirror Platform BibTeX

Zero-Shot Real-World Success

Unseen-Scene Success (from 0%)

Seed Demos Required

26.3 Hz

SuperTiny Control Freq.

SCROLL

01 — Abstract

Breaking the
Data Bottleneck

A dual-stage self-evolving framework that generates massive, high-quality robot trajectories from minimal human input — and projects them into parallel worlds for real-world transfer.

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 robot learning.

To address this, we propose Seed2Scale, a self-evolving data engine with parallel worlds expansion, operating through a dual-stage pipeline. In the self-evolution stage, a lightweight SuperTiny collector (π_small) enables robust exploration from only 4 seed trajectories, a frozen Qwen3-VL verifier (Φ_VLV) autonomously scores and filters trajectories to prevent model collapse, and a SmolVLA target model (π_target) learns exclusively from curated trajectories. In the parallel worlds stage, self-evolved trajectories are projected into diverse environments of the same task, enhancing adaptability to unseen scenes and real-world deployment.

Experiments demonstrate consistent scaling as iterations progress. The self-evolution stage lifts Seen-scene success from 22.18% to 68.57%, while Parallel Worlds raises Unseen-scene success from 0% to 65.34%. Notably, Seed2Scale achieves a 75.38% success rate in zero-shot real-world evaluations, where baseline methods fail completely (0%). Code, datasets, assets, and model weights will be open-sourced.

02 — Overview Video

Display Video

A comprehensive demonstration of the Seed2Scale pipeline — from seed data collection to self-evolution, parallel worlds expansion, and zero-shot real-world deployment.

03 — Framework

The Seed2Scale Pipeline

From 4 boundary-point demonstrations to massive high-quality training data — a dual-stage self-evolving ecosystem combining self-evolution with parallel worlds expansion.

Seed2Scale self-evolving data engine pipeline overview with parallel worlds projection

Fig. 1 — The Seed2Scale self-evolving data engine. Starting with as few as four human demonstrations, the framework utilizes a lightweight SuperTiny VLA model for large-scale parallel data collection. A pretrained Vision-Language Model acts as a Verifier, autonomously scoring and filtering the raw trajectories to prevent model collapse. A parallel-world projection mechanism then achieves visual generalization on top of the position generalization accomplished through self-evolution, yielding large-scale, high-quality, and diverse datasets for model training.

04 — Methodology

Heterogeneous Synergy + Parallel Worlds

Three specialized model roles and one trajectory-level augmentation stage, each exploiting distinct scale asymmetries to achieve what no single model can accomplish alone.

SuperTiny Collector

A lightweight 48M-parameter VLA model: a ResNet-18 vision encoder, frozen T5-Small language encoder, and a compact MLP state embedding are unified into a conditioning memory, processed by a query-based Transformer decoder with action chunking (K=30) and temporal ensembling. Achieves 26.3 Hz inference — 3.6× faster than Diffusion Policy — and bootstraps robustly from just 4 seed trajectories.

π_small · 48M params · 26.3 Hz

VLM-as-a-Verifier

A two-stage verification pipeline: a rule-based success gate filters failed trajectories by design, then a frozen Qwen3-VL-32B serves as the Vision-Language Verifier (VLV), scoring execution quality from the task instruction, rollout video, and a reference demo. Only trajectories with S_τ ≥ γ enter D_silver, preventing data-poisoning spirals.

Φ_VLV · Qwen3-VL-32B · Frozen

Parallel Worlds

A trajectory-conditioned stage that projects self-evolved trajectories into diverse target worlds of the same task — with task-relevant objects at source-aligned poses but varied assets, backgrounds, and cameras. Feasibility checks (workspace, kinematics, collision, contact, smoothness) precede re-collection, and VLV-filtered rollouts augment the final dataset.

Trajectory projection · 10 worlds · 3DGS

SmolVLA Target Model

The final policy (π_target) trained via Conditional Flow Matching on the curated dataset D_final. 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

Dual-Stage Pipeline: Self-Evolution → Parallel Worlds

4 Seed Demos

Boundary-point teleoperation

SuperTiny

Parallel rollout N envs

Gate + VLV

Rule gate + score S ≥ γ

D_silver

Curated trajectories

Parallel Worlds

Project τ into M target worlds

D_final → SmolVLA

Target model training

05 — Qualitative Results

Task Demonstrations

Direct comparison between the SeedOnly baseline and the Seed2Scale* policy (self-evolution) across representative manipulation tasks.

Kitchen Cleanup · Task 1 SEEDONLY SEED2SCALE*

SeedOnly (4 demos)

Seed2Scale* (Ours)

Can Stacking · Task 4 SEEDONLY SEED2SCALE*

SeedOnly (4 demos)

Seed2Scale* (Ours)

Seed2Scale* vs. MimicGen on GR-1 Tasks

Smoother Motion, Higher Success

Self-evolved trajectories are physically plausible and smooth. 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% (+58.50) · Replay: 67.86%

Cylinder GraspingMimicGen

Policy: 37.25% · Replay: 21.00%

Cylinder GraspingSeed2Scale*

Policy: 66.00% (+28.75) · Replay: 86.96% (4.1×)

06 — Quantitative Results

Performance Metrics

Comprehensive evaluation across simulation and real-world tasks, demonstrating consistent scaling with only 4 seed demonstrations and zero-shot transfer to unseen scenes.

Zero-shot real-world success (baselines fail completely)

0.00% → 75.38% · Pick-and-Place

Average success on unseen scenes via Parallel Worlds

0.00% → 65.34% · 3 unseen worlds

Replay success over MimicGen (Cylinder Grasping)

21.00% → 86.96%

SuperTiny speed vs. Diffusion Policy

38ms vs 136ms inference

Agibot A2: Seen vs. Unseen Scene Generalization

Multi-task joint training · 4 seed demos per task · 396 independent trials · Seen = 10 parallel worlds (data generation) · Unseen = 3 disjoint worlds (evaluation only)

Task	Evaluated on Seen Scenes			Evaluated on Unseen Scenes
Task	SeedOnly	Seed2Scale*	Seed2Scale	SeedOnly	Seed2Scale*	Seed2Scale
Kitchen Cleanup	24.63%	71.43%	68.50%	0.00%	0.00%	75.51%
Cup-to-Cup Transfer	23.50%	64.14%	60.61%	0.00%	0.00%	61.62%
Can Stacking	7.50%	65.90%	67.70%	0.00%	0.00%	69.19%
Air Fryer Manipulation	33.08%	72.82%	71.21%	0.00%	0.00%	55.05%
Average	22.18%	68.57%	67.01%	0.00%	0.00%	65.34%

Seed2Scale* = self-evolution only (spatial generalization in the source world); Seed2Scale = full engine with Parallel Worlds (visual generalization). Both SeedOnly and Seed2Scale* score 0% on unseen scenes — scene diversity, not additional action supervision, is what unlocks generalization.

GR-1 Tasks: Seed2Scale* vs. MimicGen

Metric	Task	MimicGen	Seed2Scale*
Policy Success	Cylinder Grasping	37.25%	66.00%
	Wheel Manipulation	34.75%	93.25%
	Average	36.00%	79.63%
Replay Success	Cylinder Grasping	21.00%	86.96%
	Wheel Manipulation	48.50%	67.86%
	Average	34.75%	77.41%

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 action curves comparing Expert Demonstration, MimicGen and Seed2Scale

Fig. — Action Curve Comparison. Robot action curves generated by expert demonstration, MimicGen, and Seed2Scale*. Seed2Scale* closely tracks the Expert Demonstration in both amplitude and phase, while MimicGen diverges with severe high-frequency jitter — reflecting its higher Total Variation and Jerk scores in the table.

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

Fig. — Collector Comparison over Self-Evolution Iterations. SuperTiny (full model) converges faster and reaches the highest final success rate compared with its ablated variant SuperTiny⁻ (without VLV-based quality filtering), ACT, and Diffusion Policy. The gap to SuperTiny⁻ confirms that enforcing high trajectory quality — not just task completion — is essential for robust policy learning.

Scaling: Target Model (SmolVLA) vs. Self-Evolution Iteration

Can Stacking — single-task SmolVLA checkpoints across 8 iterations

Target model (SmolVLA) success rate across 8 self-evolution iterations on Can Stacking

Success Rate — SeedOnly vs. Seed2Scale*

Per-task comparison on Agibot A2 — Seen scenes

Kitchen Cleanup

Cup-to-Cup Transfer

Can Stacking

Air Fryer Manipulation

SeedOnly Seed2Scale* (Seen)

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)
Stage 1Rule-based Success Gate
Stage 2VLM Quality Scoring
Input 1Task instruction g
Input 2Rollout video V_τ
Input 3Reference video V_ref
Output ScoreS_τ ∈ [0, 10]
Threshold γFixed a priori (rubric)

Parallel Worlds

Trajectory-Conditioned Visual Generalization → D_final

10 Worlds

Worlds (A2 suite)10 scenes · some 3DGS
Seen / Unseen10 data · 3 eval-only
World Spec. C_mAssets · poses · background · cameras
Object PosesSource-aligned across worlds

Projection M_m(τ_F)Control trajectory re-targeting
Feasibility CheckKinematics · collision · contact
Quality FilterVLV re-scoring after rollout
Real-World Setup7 sim scenes → real scene

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_final (curated)
DeploymentAgibot A2 / GR-1

SuperTiny VLA Architecture: Vision-Language-State encoding, conditioning memory, Transformer decoder with cross-attention, and Action Head with Temporal Ensembling

Fig. — SuperTiny VLA Architecture. The model integrates vision (ResNet-18), language (T5-Small), and robot state encodings into a unified conditioning memory M_t, which is processed by a lightweight Transformer decoder via cross-attention from learnable query embeddings to predict action chunks. Temporal ensembling ensures smooth and consistent control at 26.3 Hz.

08 — Multi-Task Demo

Four-Task Generalization

Multi-task joint training with cross-task knowledge transfer consistently outperforms single-task training. Evaluated on Seen scenes (10 parallel worlds) and Unseen scenes (3 disjoint worlds) on the Agibot A2 platform.

Kitchen Cleanup

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

Seen: 68.50% · Unseen: 75.51%

Cup-to-Cup Transfer

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

Seen: 60.61% · Unseen: 61.62%

Can Stacking

Stack cans from both sides into the center and ensure they are placed stably. The hardest task for the 4-seed baseline, requiring complex spatial reasoning far beyond the seed demos.

Seen: 67.70% · Unseen: 69.19%

Air Fryer Manipulation

Grasp carrots from a plate, then open the air fryer and place the food inside. Tests multi-step manipulation with articulated objects, requiring precise door-handle grasping and coordinated placement.

Seen: 71.21% · Unseen: 55.05%

09 — Zero-Shot Sim2Real

From Simulation to the Real World

Policies trained exclusively in simulation — with no real-world data collection or fine-tuning — deployed zero-shot on a Pick-and-Place task, evaluated over 65 independent trials in an unseen real-world scene.

Parallel sim scenes for training data

+ real scene for evaluation

1,716

Episodes generated (225,217 frames)

vs. 4 episodes / 720 frames (SeedOnly)

Independent real-world trials

Unseen real scene

Real-world fine-tuning steps

Pure zero-shot transfer

Configuration	Episodes	Frames	Zero-Shot Real-World Success
SeedOnly	4	720	0.00%
Seed2Scale*	142	19,748	0.00%
Seed2Scale (3 scenes)	432	60,039	27.69%
Seed2Scale (5 scenes)	720	100,065	73.85%
Seed2Scale (7 scenes)	1,716	225,217	75.38%

Success rate increases monotonically with scene diversity: from 0% in the single source world (SeedOnly and Seed2Scale*) to 27.69% with 3 scenes, 73.85% with 5 scenes, and 75.38% with 7 scenes — Parallel Worlds is the primary driver of zero-shot generalization to the real world.

Parallel worlds for simulation experiments and zero-shot real-world experiments

Fig. — Parallel Worlds Overview. (a) Simulation experiment: the original world, the 10 parallel worlds projected from it (some constructed with 3D Gaussian Splatting), and the 3 unseen worlds reserved for evaluation. (b) Zero-shot real-world experiment: 7 parallel sim worlds and the real-world evaluation scene.

Real-World Demonstrations

Generalization to Unseen Scenes and Items

All demonstrations below are zero-shot: the policy is trained exclusively on simulation data generated by Seed2Scale (4 seed demos + parallel worlds) and deployed directly on the physical robot. Evaluation uses multiple target-item styles with evaluation areas aligned to the simulation data distribution, including configurations where multiple target items appear simultaneously at different locations.

Best zero-shot real-world pick-and-place results with 7-scene parallel-world training

Best zero-shot results of parallel-world training with 7 scenes — dexterous pick-and-place directly on the physical robot.

Zero-shot real-world pick-and-place demo 1

Zero-shot real-world pick-and-place demo 2

Zero-shot real-world pick-and-place demo 3

Zero-shot real-world pick-and-place demo 4

Zero-shot real-world pick-and-place demo 5

Zero-shot real-world pick-and-place demo 6

Zero-shot real-world pick-and-place demo 7

Zero-shot real-world pick-and-place demo 8

Zero-shot real-world pick-and-place demo 9

Zero-shot rollouts on the physical robot across diverse unseen target items (varied can and bottle styles) and spatial configurations — including multiple target items placed at different locations — demonstrating real-world generalization far beyond the source world, without any real-world fine-tuning.

11 — Acknowledgements

Special Thanks

Agibot A2 robot platform for large-scale experimental validation of 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.

12 — 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 with Parallel Worlds Expansion for Scalable Robot Learning},
  author  = {Tai, Cong and Zheng, Zhaoyu and Long, Haixu and Wu, Hansheng and Wang, Jialiang and Long, Zhengbin and Xiang, Haodong and Wang, Shuyuan and Shi, Rong and Cui, Zhuo and Zhang, Shizhuang and Wang, He 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 Synergy + Parallel Worlds

SuperTiny Collector

VLM-as-a-Verifier

Parallel Worlds

SmolVLA Target Model

4 Seed Demos

SuperTiny

Gate + VLV

Dsilver

Parallel Worlds

Dfinal → SmolVLA

Task Demonstrations

Smoother Motion, Higher Success

Performance Metrics

Agibot A2: Seen vs. Unseen Scene Generalization

GR-1 Tasks: Seed2Scale* vs. MimicGen

Trajectory Quality (lower = better)

Model Specifications

SuperTiny VLA

Vision-Language Verifier

Parallel Worlds

SmolVLA Target Model

Four-Task Generalization

Kitchen Cleanup

Cup-to-Cup Transfer

Can Stacking

Air Fryer Manipulation

From Simulation to the Real World

Generalization to Unseen Scenes and Items

Full Paper

Special Thanks

AGiBot provides A2 robot model and open-source support

BibTeX

Breaking the
Data Bottleneck

D_silver

D_final → SmolVLA