Abstract
In recent years, embodied intelligence has emerged as a critical interdisciplinary domain integrating multimodal perception, large model reasoning, and intelligent decision-making, demonstrating tremendous potential to expand the boundaries of intelligence and empower problem-solving in the real world. However, due to the complexity of embodied manipulations and the diversity of task scenarios, embodied intelligence research faces severe challenges, including difficulties in high-quality data collection, large embodied model construction, and training and inference optimization. These challenges hinder the development of embodied intelligence toward larger scales, stronger generality, and broader applications. This paper first introduces the relationship between general large models and large embodied models, and discusses the key challenges of embodied intelligence in three aspects: “data-model-optimization”. Then, this paper systematically reviews the core technologies of embodied intelligence, covering three main threads—embodied data and simulation, embodied “cerebellum-cerebrum” models, and embodied training and inference—with a focus on analyzing the trends in technological development. Finally, the paper discusses open challenges and future directions, including scalable data collection via integrated virtual-real collection, general embodied reasoning and manipulation, few-shot adaptation and real-time decision-making, aiming to promote technological breakthroughs and real-world applications of embodied intelligence.
Introduction: The Challenge of Physical Agency
While Large Language Models (LLMs) have mastered digital conversation, they remain "brains in a vat"—incapable of interacting with the physical world. Embodied Intelligence (EI) aims to bridge this gap by providing AI with a physical body to perceive and act. However, moving from digital reasoning to physical action introduces three primary bottlenecks:
- Data Scarcity: Unlike the internet-scale text used for LLMs, high-quality multimodal robotic data is extremely difficult and expensive to collect.
- Model Complexity: General models struggle with the spatial and temporal nuances required for fine-grained manipulation.
- Inference Efficiency: Current Vision-Language-Action (VLA) models are often too slow (3–15 Hz) for the real-time requirements of robotics, which typically require frequencies exceeding 100 Hz.
The Data-Model-Optimization Framework
To address these bottlenecks, researchers have proposed a closed-loop framework that emphasizes the iterative expansion of datasets and model refinement. This "data flywheel" approach involves continuous data collection from both manual and simulated sources, offline pre-training, and online fine-tuning during task execution.
As shown in the framework above, the process begins with expanding embodied datasets through diverse sources—including internet data, simulation, and manual collection. This is followed by offline pre-training of policy network models. The model is then deployed for downstream task fine-tuning, where real-world interaction data is fed back into the system to virtualize real scenes and generate augmented data, completing the self-improvement loop.
Hierarchical "Cerebrum-Cerebellum" Architecture
A core technological trend in EI is the transition from monolithic "End-to-End" models to a hierarchical "Cerebrum-Cerebellum" (大小脑) Architecture. This structure mimics biological systems to balance high-level reasoning with low-level motor control.
- The Cerebrum (大脑): A high-level reasoning module (often LLM/VLM-based) that decomposes complex natural language instructions into logical sub-tasks.
- The Cerebellum (小脑): A low-level execution module that translates these sub-tasks into precise, high-frequency motor commands using policy-based learning.
This taxonomy highlights how the field is diversifying its approach across data generation (e.g., RoboGen), model architectures (e.g., RT-2, SayCan), and training paradigms like Imitation Learning (ACT, BC-Z) and Reinforcement Learning (RLPD).
Datasets and the Scaling Challenge
Embodied data is currently 2–3 orders of magnitude smaller than the text data used for models like GPT-4. To reach a "Scaling Law" moment for robotics, the community is developing specialized datasets:
- Reasoning Datasets: Tools like Matterport3D, ALFRED, and the recent OpenEQA focus on scene understanding and task planning.
- Manipulation Datasets: Efforts like Open X-Embodiment represent massive cross-platform collaborations to aggregate data across different robot morphologies.
- Simulation-to-Data: Tools like RoboGen and Eureka use LLMs to automatically generate simulation tasks and reward functions, creating a path toward infinite data generation.
Key Results and Real-Time Performance
Recent advancements have significantly improved inference speed and efficiency. While standard VLA models like RT-2 (55B) operate at roughly 3 Hz, newer hierarchical systems like Helix—which pairs a 7B "slow" cerebrum with an 80M "fast" cerebellum—can achieve 200 Hz decision frequencies. Furthermore, techniques like PD-VLA (Parallel Decoding) have demonstrated a 2.5x increase in inference frequency without requiring additional training.
Economically, the impact is substantial, with projections suggesting the embodied intelligence market in China alone could reach 5.295 billion RMB by 2025, accounting for over a quarter of the global market.
Future Directions
Despite progress, several open problems remain:
- Universal Action Representations: Developing a "common language" for actions that works across quadrupeds, humanoids, and robotic arms.
- The Sim-to-Real Gap: Improving simulation fidelity to better capture complex physical dynamics like friction and soft-body interactions.
- Few-shot Adaptation: Enabling robots to learn new tasks or navigate unfamiliar environments in minutes rather than days.
By integrating virtual-real data collection with general reasoning and high-speed execution, the field is moving toward truly autonomous agents capable of assisting in complex, real-world environments.
Summary updated