A working definition
Embodied intelligence concerns the relationship between an agent, its body and its surroundings. A useful working definition is the ability to use observations to choose actions, encounter their consequences and adjust behaviour toward a goal. Researchers use the term in different ways; it does not name one model architecture or a single test of intelligence.
The body is part of the problem. A camera's position affects what can be seen, while a gripper's shape limits what can be held. Gupta and colleagues' research on learning and evolution explores how simulated body designs can make control easier to learn. Its findings concern the environments and agents studied, rather than a universal recipe for robot design. [2]
The perception–action loop, with a cup
Consider an illustrative task: moving a cup from a table to a tray. The robot estimates the cup's position from its sensors, chooses a reachable grasp and moves its arm. That movement changes the scene. Further observations may show that the cup shifted, the grasp missed or the object is slipping. The next action should respond to the updated situation.
This loop links perception, decision, action and feedback. Success means achieving a physical outcome under constraints, rather than merely describing a plausible sequence. Timing, contact and the available sensors matter. Roy and colleagues discuss why learning for physical interaction introduces challenges beyond standard machine-learning settings, including adaptation and consequential errors. The cup example explains the idea; it is not a report of a particular system's performance. [1]
In this hypothetical example, a missed grasp should lead to a fresh observation of the cup, a revised target and another action. If the robot simply repeats a fixed motion while a camera only records video, that recording does not show that visual feedback changed the next action. A feedback controller may still use other sensors; identify which signal affects which decision. This example is editorial explanation, not an IAEI experiment.
Six questions for tracing a feedback loop
The following is an editorial reading checklist, not a certification standard or a test of general intelligence. Habitat distinguishes sensors, observations, actions, state updates and task success, providing useful terms for describing a system. [5]
- Observation: what information is available to the controller, from which sensor and at what time?
- Decision: how does the controller use that information and the task goal to choose its next command?
- Action: what command reaches the body or simulator, and what motion or interaction does it produce?
- Environment change: what actually changes after the action, including the agent’s own state?
- Feedback: which new observation influences a later command? Recording a signal alone is not enough to demonstrate its use.
- Success condition: what observable result counts as completion, and when is an attempt stopped or marked unsuccessful?
These are six questions, not six required modules or sequential steps. The success rule evaluates the attempt; feedback may alter an action without updating model parameters.
Closing this loop does not establish that the system learns or generalizes. Habitat also supports agents that use classical planning without learning. Describe the controller and the tested conditions separately from claims about intelligence. [6]
Robots, language models and simulation
A robot is a physical machine; the label alone does not establish flexible learning or general intelligence. A text-only assistant can explain how to lift a cup without sensing or controlling that cup. A language model may nevertheless be one component of an embodied system. RT-2, for example, studies how visual and language representations can be connected to robotic action using both web and robot data. [3]
Embodied AI research also includes simulated agents. Habitat provides virtual environments for training and evaluating agents on tasks such as navigation. Simulation enables repeatable experiments, but a result inside a simulator does not by itself establish performance on real hardware. Researchers must examine which observations, actions and environmental assumptions carry over. [4]
What simulation, hardware and video can establish
These evidence types answer different questions. Roy and colleagues discuss the distinction between demonstrations and systematic evaluation, including the gap between simulated and physical interaction. [1]
- Simulation: useful for repeatable tests under stated sensor, physics and task settings. It does not alone establish transfer to hardware. Habitat explains why simulation helps control and repeat experiments. [6]
- Hardware trials: report the robot, environment, number of attempts, success rule and human interventions. Results support the conditions tested; a new object, surface or camera setup requires further evidence.
- Demonstration video: shows an example trajectory. Without a stated trial set and failure record, it cannot establish a success rate. Check whether feedback, resets, remote control or manual recovery are disclosed.
As a reading practice, record an unspecified item as “not reported in the material reviewed”, rather than assuming it was absent from the experiment.
What to check in a research claim
Read a demonstration as evidence about the conditions actually tested. A useful reading checklist is:
- What body, sensors, actions and task were available?
- Was the result measured in simulation or on a physical robot?
- What changed between training and evaluation, and what remained familiar?
- How many trials were reported, and were failures or human interventions described?
Keep the reported result separate from your interpretation. A promising method can still have a narrow operating range. Neither a fluent explanation nor a successful demonstration alone establishes broad physical competence. Follow the original paper for experimental details and limits.