How Machine Behavior and Embodied Intelligence Shape the Future of Autonomous Driving

On April 25, the CPC Politburo held its 20th collective study on strengthening artificial intelligence development and regulation. Academician Zheng Naning from Xi'an Jiaotong University gave a lecture and offered work suggestions.

Machine Behavior and Embodied Intelligence

We begin with a simple intersection scenario involving pedestrians, non‑motorized vehicles, and motor vehicles. Although traffic scenes are unpredictable, each object’s intuitive judgment and understanding of mutual behavior form a stable, interrelated system.

Humans quickly comprehend spatial and behavioral relationships in such scenes; autonomous driving must similarly abstract and represent these relationships to make accurate decisions. Encoding all possible dynamic changes in advance is impossible, so research must focus on adaptive behavior of multiple autonomous agents.

In F1 races, rapid tire changes illustrate collaborative robot tasks, prompting the question of how robot clusters can accomplish tasks with scientific explanations.

Discussion 1: Machine Behavior Imitation and Explanation

Explaining behavior is harder than generating it because most human actions are learned from the environment. A Turing‑machine‑like system can mimic behavior without true intelligence; explanation requires clear generalizations linked to universal principles, reflecting cognitive processes.

AI development can be divided into four stages:

Expert learning systems: encode domain knowledge and rules.

Feature engineering: provide predefined features and labels for learning.

Deep learning: feed raw data and labels to deep neural networks, achieving breakthroughs in speech and image recognition.

General AI (the fourth stage): give machines tasks and goals so they can perceive, understand, and learn like humans, aiming for consciousness and adaptability across diverse tasks.

General AI would possess self‑awareness, autonomous reasoning, planning, problem‑solving, and the ability to adapt to novel situations, requiring extensive background knowledge and common sense.

Discussion 2: Challenges Facing Machine Behavior

Beyond technical hurdles, general AI confronts ethical, social, and legal issues. Two fundamental problems arise in complex, uncertain environments:

Condition problem – it is impossible to enumerate all preconditions for a behavior.

Branching problem – it is impossible to list all latent outcomes of a behavior.

Traditional AI, based on deductive logic and formal semantics, cannot model every object or action.

A classic child‑assistance experiment shows a 1½‑year‑old spontaneously helping an adult open a cabinet, challenging AI to achieve similar spontaneous, cooperative intelligence.

Discussion 3: Scope of Machine Behavior Research

Machine behavior research focuses on intelligent machines, not traditional mechanics. It covers behavior generation, experience‑based action, explainability, and context‑driven responses.

Discussion 4: Embodied Intelligence and Behavior Generation

Embodied intelligence enables machines to autonomously perceive, learn, and act within their environment, mirroring biological evolution where intelligence arises from body‑environment interaction.

Non‑embodied learning relies on large‑scale pre‑training and fine‑tuning, independent of hardware. Embodied learning combines virtual pre‑training with reinforcement learning in specific hardware contexts.

Discussion 5: Representation Learning and Causal Reasoning for Embodied AI

Achieving human‑like cognition requires building event models that represent objects, events, and facts, enabling continual learning and optimal task strategies through perception, prior knowledge, representation learning, and knowledge bases.

Discussion 6: Human‑Machine Collaboration in Dynamic Open Environments

To make embodied intelligence more human‑like, it must be reinforced with human‑machine collaboration in open, dynamic settings. Traditional reinforcement learning excludes humans, limiting adaptability.

Human‑in‑the‑loop decision making, visual learning, imitation, and interactive training can guide autonomous systems toward more robust behavior.

Autonomous Driving: Challenges and Behavior Generation

Autonomous driving exemplifies embodied intelligence in open environments. Key challenges include:

Comprehensive perception under all weather and lighting conditions.

Understanding pre‑behaviors to infer driver intent.

Handling unexpected encounters without exhaustive rule encoding.

Ensuring cybersecurity against software vulnerabilities and attacks.

Behavior generation involves combining experience, common sense, scene understanding, and traffic situation assessment to pre‑train models, sample target states, generate candidate motion paths, select optimal trajectories, and execute safe driving actions.

Complex, dynamic traffic scenarios cannot be fully modeled; instead, autonomous systems must abstract the environment into a feasible “drivable” state space.

Cognitive Mapping for Autonomous Driving

Building a cognitive map requires representing drivable areas, traffic signs, obstacles, and pedestrians, enriched with learned attention mechanisms and intent attributes. Recursive networks integrate perception and prior knowledge into a temporal visual map, which a value‑iteration model translates into actionable driving decisions.

Simulation Testing

Before large‑scale deployment, autonomous vehicles need extensive testing—approximately 440 million kilometers to match human fatality rates. Real‑world testing would take years; simulation offers efficient, low‑cost evaluation, especially for rare abnormal traffic scenarios generated via graphics and computer vision techniques.

A large‑model‑based simulation framework can generate diverse traffic scenes to assess safety, comfort, coordination, and legal compliance, comprising datasets of real sensor data, scene descriptions, classification, and representative scenario generation.

Thank you for your attention.