Proprioception

Proprioception is a robot's sense of itself. Borrowed from the biological term for how animals know the position of their limbs without looking, it describes the internal state signals that tell a robot how its own body is currently configured. This includes the angle of each joint, how fast the joints are moving, the forces or torques being applied, where the end-effector is in space, and whether the gripper is open or closed. Unlike a camera, which senses the outside world, proprioception senses the machine itself.

This distinction is important because acting in the world requires knowing both what is out there and where your own body is. A robot reaching for an object needs to combine what it sees with an accurate sense of where its hand currently is, otherwise it cannot close the gap between the two. Proprioceptive state is therefore almost always fused with visual input to control a robot, and it is one of the things that makes physical AI different from tasks that work purely on images. There is no proprioception in a static photograph, but there is no capable robot without it.

Proprioceptive signals come from sensors built into the robot itself, such as encoders that measure joint angles and sensors that measure force or torque. These readings are sampled continuously as time-series data and combined with external perception to form the full observation a controller uses. When a robot decides how to move, it reasons over both its proprioceptive state and what it senses of the environment, because an action only makes sense relative to the body's current configuration. Keeping proprioceptive and visual streams aligned in time is part of what lets a controller coordinate the two into precise movement.

Proprioception matters because manipulation and locomotion are impossible without an accurate sense of one's own body, which is why it is a core part of the data physical AI systems learn from. For anyone building or curating robot data, it is a reminder that observations are not only about the outside world, since the robot's internal state is an equally essential input. This self-sensing is one of the clearest ways in which controlling a physical body differs from working with images or text alone.

Vision senses the external world, while proprioception senses the robot's own body, such as its joint angles and gripper state. Vision tells the robot what is around it, and proprioception tells it how it is currently configured. Both are usually needed together.

To act on an object, a robot needs to know both where the object is and where its own hand is, then close the gap between them. Vision provides the former and proprioception the latter, so fusing them is what makes precise, coordinated movement possible.

No. Purely image-based tasks have no notion of a body sensing itself, so proprioception is specific to embodied systems. This is one of the ways physical AI differs from working with static images or text.

Embodiment, Degrees of freedom (DoF), End-effector, Sensor fusion, Time-series sensor data.