Latest Breakthrough in World Models and Embodied AI: 90% Synthetic Data, VLA Performance Soars 300% | Open Source

Honghao Wang

02 Dec 2025 — 4 min read

🚀 VLA Model Performance Soars by 300%

90% of Training Data Now Generated by a World Model — for the First Time

A domestic world model developer has achieved a breakthrough in embodied intelligence — and has open-sourced all model code and training frameworks.

---

The Data Bottleneck in Embodied Intelligence

The biggest challenge for deploying embodied intelligence in open-world environments has not been algorithms, but the extreme scarcity of high-quality, large-scale, real robot interaction data.

Key issues with real-machine data collection:

High cost and significant time investment
Difficulty covering diverse open-world scenarios
Severe limits on scalability and generalization

While simulation can quickly produce data, it introduces a Sim-to-Real gap, making real-world deployment less robust.

---

Why World Models Matter

World models learn the rules and patterns of the physical world, enabling them to generate high-fidelity, controllable, and diverse embodied interaction data.

This helps overcome the shortage of real-machine data.

Breakthrough:

Chinese world model company GigaVerse, backed by Huawei, has released GigaWorld-0, increasing the proportion of world model–generated data in VLA training to 90%.

---

Performance Gains

The trained VLA model showed ~300% improvement across three generalization dimensions:

New Textures — unseen surface materials during training
New Viewpoints — unseen observation angles during training
New Object Positions — unseen spatial layouts during training

This marks a new stage: data-efficient, highly generalizable, and low-cost embodied intelligence.

---

GigaWorld-0 Architecture

GigaWorld-0 comprises two coordinated components:

GigaWorld-0-Video — generates richly textured, visually realistic embodied operation data using video generation foundation models.
GigaWorld-0-3D — combines 3D generation, 3D Gaussian Splatting reconstruction, and differentiable physics for accurate geometry and dynamics.

GigaWorld-0-Video Improvements

Addresses low computational efficiency and poor fine-detail control by boosting:

Sparse Attention modeling capacity — memory- and latency-efficient long-range spatio-temporal attention.
Dynamic Expert Computation capability — increases diversity and controllability.

Sparse Attention:

Uses a Diffusion Transformer (DiT) with sparse attention — local spatio-temporal neighborhoods and key semantic regions only — reducing complexity and cost.

Mixture-of-Experts (MoE):

Integrates MoE within DiT feedforward layers, routing video tokens to specialized experts for fine-grained control inspired by DeepSeek V3.

---

GigaWorld-0-3D: Generative + Reconstruction

Enhances scene modeling under sparse observations with differentiable physics simulation for robotic manipulation.

Capabilities:

Geometrically consistent & visually realistic static backgrounds
Accurate modeling of robot–object dynamics

Generative Reconstruction

Initialize Gaussian scene representation from sparse views
Apply view-repair generation to fix distortions
Reconstruct high-precision 3DGS for novel view synthesis fidelity

---

Differentiable Physics Engine

Based on Physics-Informed Neural Networks (PINNs):

Generate trajectories with random physical parameters
Train differentiable surrogate models for system dynamics
Optimize parameters via gradient descent to match real motion

Produces physically plausible and interaction-reliable data.

---

FP8 Precision + GigaTrain Framework

Industry first: world model trained end-to-end in FP8 precision — optimized for energy efficiency.

Combines FP8 + Sparse Attention to balance visual fidelity & compute cost
Powered by GigaTrain: open-source distributed training framework supporting:
DeepSpeed ZeRO
FSDP2
FP8 mixed precision
Gradient checkpointing

Repo: GitHub

---

Benchmark Results

PBench (Robot Set) comparison:

Models tested: Cosmos-Predict2-14B, Cosmos-Predict2.5-2B, Wan2.2-5B, Wan2.2-14B

GigaWorld-0 — smallest model at 2B parameters
Achieved highest overall score and outperformed in key metrics

---

Real-World Impact

GigaWorld-0 is a generalized embodied data engine:

Validated on GigaBrain-0 VLA model across:
New texture generalization
New viewpoint generalization
New object position generalization

Increasing synthetic data proportion steadily improved task success rates and motion accuracy on real robots.

---

Jija Vision: Leading in World Models & Embodied Brains

Founded in 2023 as China’s first physics AI company, focusing on World Model Platforms × Embodied Foundation Models.

Leadership Team

Huang Guan, CEO — PhD in AI, Tsinghua University; ex-Horizon Robotics; experience at Samsung China Research & Microsoft Research Asia.
Zhu Zheng, Chief Scientist — PhD, Chinese Academy of Sciences; postdoc at Tsinghua University; >17,000 citations, h-index 50.
Top researchers from Tsinghua, Peking Univ, CAS, USTC, WashU, CMU, and leaders from Microsoft, Samsung, Baidu, Bosch, etc.

---

Industry Engagements

World-class expertise in World Models and Embodied Brains
Contracts with leading autonomous driving OEMs
Applications spanning research, education, exhibitions, industry services, household robotics
Nov. 2023: Completed 100M RMB A1 financing — Huawei Hubble & HC Capital

---

AiToEarn官网 — open-source platform for AI content generation, publishing, and monetization across:

Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X (Twitter)

Integrates:

AI content creation
Cross-platform publishing
Analytics & model ranking (AI模型排名)

Open-source repo: GitHub

---

If you like, I can also create an infographic-friendly condensed version of this Markdown so your readers can quickly grasp the structure and key achievements of GigaWorld‑0. Would you like me to do that?