Embodied AI

Embodied AI connects computation to a system that can sense and act in a physical environment. That connection changes what counts as an action: a digital system can return a label or instruction, while a robot must turn a selected action into commands that affect a body and its surroundings. The body and environment matter because different machines have different movement capabilities, and physical interactions can change what the system senses next.

A useful distinction is between a high-level intention, such as moving an object, and the lower-level movements that carry it out. A model can contribute to choosing or representing an action, but physical control also depends on the robot and its controllers. When reading a robotics paper, identify the system’s inputs, its action outputs, and the physical platform on which those outputs are used.

• RT-2: New model translates vision and language into action

  An example of a model representing robot actions as outputs that can be used for control.

Embodied AI

AI that perceives and acts within a physical environment.

Vision-language-action model

A model that uses visual and language information to produce actions for controlling a robot.

A robot plans from information about its surroundings, but that information can be incomplete or uncertain. MIT’s robotics teaching materials use mapping an unknown space, estimating a robot’s location in a known space, and plotting a route through a maze as linked planning problems. These examples help distinguish perception from planning: sensing supplies measurements, estimation uses them to infer a useful state, and planning selects a route or action toward a goal.

For an embodied system, a plan is tied to what the robot can sense and do. Ask what the sensors observe, what the system must infer, and which choices remain open when the observations are uncertain. A route selected from a map is still a plan; it does not itself guarantee the robot has moved safely through the physical space.

• Introduction to Robotics

  Use the overview to connect sensors and motion planning to robot mechanisms and control.

Control connects a desired movement to commands that a robot’s actuators can carry out. The robot’s mechanics and dynamics shape that connection: a plan that looks sensible at an abstract level may need to be adjusted for the machine’s movement and interaction with its surroundings. Feedback adds another step. The system senses what happened, compares it with what was intended, and can adapt its actions when the scene changes or an object slips.

This distinction is useful when evaluating claims about a robot. Ask whether a result concerns interpreting an instruction, planning a movement, issuing motor commands, or completing the physical task. MIT’s robotics courses treat motion planning, control, actuators, and sensors as related but distinct topics. Keep the robot’s actual body and control setup in view when interpreting an action model’s output.

• Underactuated Robotics

  For a deeper look at nonlinear dynamics, motion planning, and control in mechanical systems.

• Gemini Robotics brings AI into the physical world

  An example of a system monitoring surroundings and adjusting actions when the situation changes.