Five Levels

mlstack 2 min read

Understanding Progress in Embodied AI: The Five Levels of Embodied AGI

This page provides a general overview of a useful framework for thinking about progress toward advanced physical AI systems. The five-level taxonomy was proposed by researchers Yequan Wang and Aixin Sun in their 2025 paper "Toward Embodied AGI: A Review of Embodied AI and the Road Ahead".

The framework is inspired by the familiar levels used for autonomous driving. It offers a simple way to understand how AI-powered robots could evolve from performing basic repetitive jobs to handling complex, open-ended tasks in the real world.

Level 1: Single-task completion

Basic assistance with one specific, limited task (for example, simple pick-and-place operations or vacuuming) in controlled environments. Very little flexibility if conditions change.

Level 2: Compositional task completion

Can combine simple actions to achieve goal-directed tasks, such as “clean the table” by picking up and moving multiple objects. More flexible than Level 1, but still limited to familiar combinations.

Level 3: Conditional general-purpose task completion

Handles several types of tasks with some ability to adapt to changes or new instructions. Uses more types of sensors and usually works with human supervision or monitoring.

Level 4: Highly general-purpose robots

Performs a wide variety of tasks across different environments with good skills, planning, and communication. Needs only occasional help in most everyday situations.

Level 5: All-purpose robots

Fully independent agents that can safely and adaptively handle diverse, novel real-world tasks — approaching human-level physical intelligence in everyday environments.

Why This Framework Is Useful

Each higher level demands improvements in sensors, planning ability, quick reactions to surprises, and the capacity to work in new situations. Today, most robots operate at early levels (L1 or L2). Advancing toward higher levels could eventually bring versatile helper robots into homes, factories, healthcare, and exploration.