Tsinghua survey reframes embodied AI as agents: L0-L5 autonomy levels

新智元 · wechat · 2026-11-13

A Tsinghua team released a survey on Agentic Embodied Intelligence (AEI), reframing embodied AI research from an agent perspective (project page: agenticei.github.io/AEI-Survey-page).

Trajectory: from feedback control and imitation/RL skill learning, through deep visuomotor policies and foundation-model-driven generalist robot policies, to task-level autonomous organization.
Autonomy levels: an L0-L5 framework based on human-machine responsibility allocation—continuous teleoperation (L0) up to conditional autonomy (L3), sustained in-domain task execution with failure recovery (L4), and open-world general autonomy (L5); recognizing capability boundaries and safely asking for help counts as autonomy.
Online cognitive loop: four coupled model functions—perception/embodied understanding, reasoning/planning, action execution, and closed-loop evolution—with distinct evidence standards for control tweaks vs. persistent memory vs. system-wide changes.
System support: cloud-edge-device real-time execution (sporadic long latencies can make commands act on stale state), safety spanning perception-planning-control, an Agentic Embodied Experience representation, and evaluation protocols organized around five cognitive demands.

Original post →

Tsinghua survey reframes embodied AI as agents: L0-L5 autonomy levels

More from Embodied