zte-terminators/seed2scale-example-data

Seed2Scale Multi-GPU Data Generation Results

(Disclaimer: Due to capacity limitations, this warehouse only provides partial trajectory examples. For complete examples or business cooperation, please contact the corresponding author: shen.tao5@zte.com.cn)

This folder presents a real multi-GPU Seed2Scale data generation experiment conducted on a single workstation. The results summarize both data generation throughput and trajectory quality under different GPU configurations, with all reported trajectories generated within 24 hours per configuration.

The experiment is defined by the following task setup:

• Scene: Kitchen Scene
• Embodiment / Policy: AgiBotA2 · SmolVLA
• Instruction: Pick up the red potato chip and place it into the yellow basket
• Generation Window: 24 hours per configuration

In addition to aggregated metrics, this folder also includes a trajectory replay demo together with visual summaries of scaling efficiency, collection accuracy, data volume, per-GPU load balance, and spatial sampling coverage.

Visual Demo

The following GIF provides a random visualization of the collected robot trajectories. It illustrates the end-effector motion across multiple sampled positions during automatic data generation.

Experiment Overview

This benchmark evaluates how Seed2Scale behaves as compute resources scale from 1 GPU to 2 GPU and 4 GPU. The main questions are:

• How much additional data can be generated as GPU count increases?
• Whether trajectory success rate remains stable under larger parallel workloads.
• Whether the 4-GPU run maintains balanced task allocation across devices.
• Whether spatial sampling coverage remains consistent across configurations.

Hardware Configuration

Component	Model	Quantity
CPU	AMD Ryzen Threadripper PRO 7975WX 32C/64T, 4.0–5.3 GHz	1
Motherboard	ASUS Pro WS WRX90E-SAGE SE Workstation Motherboard	1
Memory	Samsung 64GB DDR5 RECC 4800	4
Graphics Card	Inno3D RTX 5090D V2 24G Triple-Fan Graphics Card	4
SSD	ZhiTai TI600 2TB NVMe M.2 SSD PCIe 4.0	1
RAID Controller	ASUS PIKE II 3108-8i 240PD 2G RAID Controller	1
CPU Cooler	AMD SP6 360 All-in-One Liquid CPU Cooler	1
Power Supply	Great Wall 2400W Platinum 80PLUS Full Modular Power Supply (Dual PSU Coupling)	2
Case	Dual-Tower Workstation Case	1

KPI Summary

Configuration	Successful Trajectories	Failed Rollouts	Success Rate	Data Size	Video Duration
1 GPU	17,480	307	98.27%	731.51 GB	33 h 2 m
2 GPU	31,551	566	98.24%	1,321.51 GB	59 h 37 m
4 GPU	52,940	904	98.32%	2,133.89 GB	100 h 2 m
Total / Combined	101,971	1,777	98.29%	4,186.91 GB	192.7 h

Quantitative Results

1. Scaling Performance

This figure compares actual trajectory throughput against ideal linear scaling. The 2-GPU configuration reaches 90.3% of ideal scaling efficiency, while the 4-GPU configuration reaches 75.7%. This indicates strong scaling benefits with a moderate efficiency drop caused by multi-instance coordination overhead.

2. Collection Accuracy and Success Rate Stability

This figure shows successful and failed rollout counts together with success rate. Across all configurations, success remains highly stable within 98.17%–98.40%, indicating that larger-scale generation does not materially reduce task success quality.

3. Data Volume Comparison

This figure compares total generated data size, video duration, and trajectory count across 1 GPU, 2 GPU, and 4 GPU settings. Relative to 1 GPU, the 4-GPU configuration achieves 3.03× more successful trajectories, together with substantial gains in total storage volume and video duration.

4. Comprehensive Metrics Comparison

This figure provides a compact side-by-side summary of the principal experiment metrics, including trajectories, success rate, failed rollouts, storage footprint, and total video duration.

5. Per-GPU Load Balance in the 4-GPU Run

This figure verifies distributed scheduling quality in the 4-GPU run. The four GPUs exhibit nearly identical workload allocation, and per-GPU success rates remain tightly clustered, demonstrating stable balancing behavior across devices.

Spatial Sampling Distribution

The following figures visualize end-effector sampling locations from evaluation_data.csv. Blue points indicate successful rollouts and red points indicate failed rollouts. The overall spatial coverage pattern remains highly consistent across GPU configurations, suggesting that scaling compute does not noticeably distort the data distribution.

1 GPU	2 GPU	4 GPU

Summary

Overall, the results show that Seed2Scale can produce large-scale robotic manipulation data efficiently on a single workstation while maintaining high success rates and stable spatial coverage. The 4-GPU setting delivers the largest data volume and best throughput, while preserving strong task accuracy and balanced per-GPU execution.