Beating NVIDIA with Just a Few Samples: China’s First Ultra-Low-Sample Embodied Model Debuts and Wins Top Conference Championship

# Bridging the Gap Between Vision-Language and Robotic Manipulation

**Date:** 2025-10-16 12:49 Beijing  

![image](https://blog.aitoearn.ai/content/images/2025/10/img_001-228.jpg)  

---

> **China's first domestic few-shot, general embodied manipulation foundation model released — connecting vision-language understanding with robotic control.**

---

## The Data Bottleneck in Embodied Intelligence

In **natural language** and **computer vision** research, large datasets are common.  
In **embodied intelligence**, however, high-quality data is scarce.  

**Why?**

- Real-world robotic manipulation involves **complex physical interactions**, **real-time feedback loops**, and **dynamic environments**.  
- Data collection is **costly**, **time-intensive**, and **hard to scale**.  
- Datasets with hundreds of thousands of physical interaction samples are rare.

---

## Limitations in Current Vision-Language-Action (VLA) Models

Even with strong semantic capabilities, existing VLA models:

- **Depend heavily** on large annotated datasets.
- Show **poor generalization** in real manipulation tasks.
- Struggle with **cross-scene adaptation** when data is limited.

**The challenge:**  
Enable embodied robots to:

1. Learn **quickly** in few-shot settings.
2. Execute **accurately** in the real world.
3. Transfer learned skills **flexibly** across scenarios.

---

## Introducing FAM-1

**FiveAges** (中科第五纪), a Chinese embodied intelligence startup, announced **FiveAges Manipulator-1 (FAM-1)**.

Core foundation:

- Based on NeurIPS 2025 paper — *BridgeVLA: Bridging the Gap between Large Vision-Language Models and 3D Robotic Manipulation*.
- First effective integration of **large-scale VLM knowledge transfer** with **3D robotic control spatial modeling**.

**Performance Highlights:**

- Requires **only 3–5 samples per task** for training.
- Achieves **97% success rate**.
- Outperforms state-of-the-art (SOTA) baselines.
- Winner of **CVPR 2025 Embodied Manipulation Challenge**.

---

## FAM-1: From VLA to BridgeVLA

**Goals:**

- Address scarcity of manipulation data.
- Improve generalization **across tasks and environments**.

**Key Innovations:**

1. **Multi-type Data Integration**  
   - Builds a multi-dimensional manipulation knowledge base.  
   - Secondary pretraining to mine implicit VLM knowledge.  
   - Improves goal/scenario understanding and generalization.

2. **3D Heatmap Alignment**  
   - Aligns outputs of VLM with inputs of VLA in the spatial domain.  
   - Effective fine-tuning with only **3–5 samples per task**.  
   - Improves spatial comprehension and data efficiency.

---

## Architecture Overview

FAM-1 consists of **two core modules**:

1. **Knowledge-driven Pretraining (KP)**  
   - Uses web-collected large-scale images/videos.
   - Builds an **operation-oriented** knowledge base.
   - Second-stage pretraining to adapt VLM to manipulation tasks.
   - Predicts robotic arm key positions and trajectories.

2. **3D Few-Shot Fine-Tuning**  
   - Avoids the “dimensional bottleneck” of flattening 3D data to 1D vectors.
   - Uses **3D heatmaps** to preserve spatial structure.
   - Reduces sample requirements drastically.

---

## Main Experimental Results

### Benchmarks: RLBench and Colosseum

- Compared with Microsoft, MIT, Stanford, and other teams.
- **RLBench success rate:** 88.2% (6% higher than SOTA).
- Major lead in tasks:
  - Insert Peg
  - Open Drawer
  - Sort Shape
  - Door Close
  - Hammer Strike
- Average success improvement: **30%+**.

![image](images/img_002.jpg)

---

## Real-World Deployment

### Few-Shot Superiority

- Compared against:
  - RVT-2 (NVIDIA)
  - PI0 (Physical Intelligence)
  - SpatialVLA (Shanghai AI Lab)
- **Only 3–5 samples** per basic task:
  - Achieves **97% success rate**.
  - Shows robustness to distractors, lighting changes, and background variation.

![image](https://blog.aitoearn.ai/content/images/2025/10/img_003-200.jpg)

---

## Summary and Outlook

FAM-1 delivers:

- **Few-shot learning** for real-world robotic arm manipulation.
- Cross-scenario generalization via **implicit multimodal knowledge transfer**.
- Preservation of **3D spatial structure** for precise control.

**Future Directions:**

1. **Enhance** generalization and reliability in operational environments.
2. **Promote industrial applications** of foundational models.
3. **Expand** foundational models to navigation tasks.

**Related Work:**  
- *EC-Flow* — Self-supervised learning of manipulation strategies from **action-unlabeled videos**.  
- Accepted at **ICCV 2025**.  
- Robots can learn manipulation simply by watching human operation videos — reducing application barriers.

---

## In the Larger AI Ecosystem

Platforms like [AiToEarn官网](https://aitoearn.ai/) empower creators and researchers to:

- Generate AI content.
- Publish across multiple platforms.
- Analyze performance.
- Access [AI模型排名](https://rank.aitoearn.ai) for model comparison.

This parallels how **FAM-1** unifies knowledge transfer and real-world fine-tuning, enabling wide-scale sharing and monetization of technical innovations — whether in robotics or content creation — across networks like Douyin, WeChat, Bilibili, Facebook, and X (Twitter).

---

[Read Original](2650995884)