A household robot may understand every word in “prepare a banana milkshake” and still fail to do what its user intended. It has to choose suitable ingredients and tools, work out a feasible sequence, cope with a missing item or broken appliance, and respect the user’s health, safety and preferences. The prompt names the task. It does not necessarily spell out the purpose, assumptions or limits that should guide each step.
Autonomous vehicles, drones and other systems face the same problem. A plan that makes sense when it is generated may no longer fit the situation once conditions change. If the system follows it blindly, it may complete the stated task while losing sight of the reason for doing it.
A Review published online August 6, 2026, in National Science Review takes this gap as its starting point. In “Intent-Driven Embodied Artificial Intelligence,” Nanning Zheng of the Institute of Artificial Intelligence and Robotics at Xi’an Jiaotong University proposes IDEAI, a system-level framework that organizes embodied AI around intent rather than task execution alone. Here, intent is the commitment that ties together what a person wants, what the environment allows, what the system knows and what it eventually does.
Why prompts and rewards are not enough
Embodied AI is broader than robotics. It covers systems that perceive, reason and act through continuing contact with the physical world, including autonomous vehicles, drones and other physically situated agents. Large language models, multimodal learning, reinforcement learning and vision–language–action models have made these systems better at navigation, manipulation and interaction. Many, however, still depend on task-specific rewards, predefined objectives or prompt-level instructions.
Those mechanisms work best when the goal is clear and the world stays stable. In open environments, neither is guaranteed: information is incomplete, circumstances change and requirements may be revised while the system is acting. A local action can succeed even as the overall plan drifts. Over time, the agent may rely on an assumption that is no longer true or find an efficient solution that violates a physical, social or safety constraint.
In IDEAI, intent has a precise technical meaning. It is not a synonym for a task, target state, prompt or reward signal. It is a structured commitment that can be inspected, revised and verified, combining a desired future state with assumptions about the current situation, expected consequences, execution constraints and value boundaries. A task says what to do. Intent also records why, under what conditions and within what limits.
Once intent is explicit, the system has something it can check. It can test whether a plan still serves the original purpose, whether an instruction is ambiguous or whether a key assumption has failed. It can then divide broad goals into sub-intents, pause an unsafe action or replan when new evidence changes the situation.
How intent enters the system
In the milkshake example, semantic grounding ties the instruction to actual ingredients, tools and possible actions. Concept generation and learning supply reusable knowledge about objects, skills and likely consequences. Intent modeling turns the request into commitments that can be broken down and revised. Value alignment brings safety, legality, human preferences and ethical or social considerations into the decision before a plan is chosen. These four layers form the core of IDEAI.
They do not run in a straight line. Perception changes the concepts and plans available; intent directs attention and action; value alignment rules some plans in or out. Action then changes the environment and produces new evidence, which may force the system to update its knowledge, its plan or even its intent.
To show how this could be built, the Review groups existing work into seven technical pillars: intent capture and semantic modeling; human priors and intuitive reasoning; world models and contextualized planning; conflict detection and multi-agent coordination; bidirectional alignment and explainable interaction; safety boundaries and verifiability; and online learning and knowledge evolution. The list is not a recipe for a single architecture. It maps the handoffs where intent can be distorted—when a command is misread, an assumption proves false, commitments conflict, a plan is unsafe or knowledge is out of date.
The paper’s main theoretical contribution lies not in proposing another algorithm, but in organizing embodied AI around explicit intent. Its four layers show how intent connects perception, knowledge, planning, action and values, while its seven pillars identify the interfaces at which those connections can be established, checked and revised.
“For embodied AI, completing the task is only part of the job,” Zheng said. “A system also needs to keep track of why it is acting, notice when its assumptions no longer hold and adjust without crossing safety or value boundaries.”
Task graphs, goal-conditioned policies, hierarchical learning, world models, human feedback, safety barriers, runtime monitors and formal verification can all play a part. IDEAI provides the organizing structure for keeping the commitments that guide action grounded, inspectable, value-constrained and open to revision from instruction through execution.
Where explicit intent may help—and where it may not
The Review looks at domestic service robots, autonomous-driving systems, low-altitude drones, emergency-rescue swarms and heterogeneous robot teams for lunar construction. These are settings in which goals are rarely fully specified in advance, plans unfold over time and mistakes can have serious consequences. An explicit account of intent could help separate negotiable preferences from hard safety limits—and clarify when the system can act on its own, when it should ask for clarification and when it needs human approval.
That extra structure is not free. It requires memory and computation, and mistakes in grounding, concept formation or value judgment can propagate through the system. Intent must also be stable enough to support a long plan without becoming so rigid that the system ignores new evidence. The right balance between stability and adaptability remains an open problem.
Nor is explicit intent needed everywhere. For short, well-defined tasks, low-level motor control or stable environments, conventional feedback control or implicit objectives may be simpler and more effective. IDEAI is most relevant when instructions are incomplete, decisions unfold over time, agents must coordinate or mistakes can cause serious harm.
IDEAI shifts the question asked of embodied AI. Finishing the job is no longer the whole test. The system must also keep the purpose of the job in view as the world changes around it.
The work was supported by the Fundamental and Interdisciplinary Disciplines Breakthrough Plan of the Ministry of Education of China under grant JYB2025XDXM504.
About the Institute of Artificial Intelligence and Robotics at Xi'an Jiaotong University
Founded in 1986, the Institute of Artificial Intelligence and Robotics at Xi’an Jiaotong University is one of the earliest research institutions in China dedicated to artificial intelligence. Its interdisciplinary research spans artificial intelligence, computer vision and pattern recognition, robotics, machine perception, intelligent systems, and human–machine interaction. Drawing on national research platforms, including the National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, the Institute conducts fundamental research and drives technological innovation in intelligent systems capable of perceiving, learning, reasoning, and interacting in complex real-world environments.
National Science Review
Literature review