Proprioception

A robot's sense of its own joint positions, velocities, forces, and posture, as opposed to sensing the outside world.

Proprioception was originally a physiology term for the sense of one's own limb position, movement, and effort; the British physiologist Charles Sherrington used the word in his writing in 1906, contrasting it with exteroception, the sensing of the outside world. In robotics it refers to measurement of the robot's own state: encoders give joint angle and velocity, motor current or torque sensors give force, an IMU gives body attitude and angular velocity, and gripper opening width is often counted too. This data is low-dimensional, low-noise, and high-frequency, so nearly every policy uses it — for example, π0's observation consists of multiple camera images, a language instruction, and a proprioceptive state made of joint angles, with the proprioceptive state projected through a linear layer and fed into the model alongside the other tokens. Legged locomotion that relies only on proprioception, with no camera at all, is called blind locomotion.

ExampleIn Unitree's open-source unitree_rl_gym, the G1 walking policy's observation includes IMU angular velocity, projected gravity, each joint's angle relative to its default pose, and joint velocity — all proprioception — concatenated with a velocity command, the previous action, and the gait phase. Body linear velocity is hard to measure accurately on the real robot, so it appears only in the privileged observation used for simulation training.

Also called

Proprioceptive State, Robot Self-State Sensing

Related

Exteroception · Rotary Encoder · State / Proprioception Encoder · State Estimation · Blind Locomotion · Inertial Measurement Unit

Sources

Wikipedia: Proprioception
π0: A Vision-Language-Action Flow Model for General Robot Control (arXiv 2410.24164)
unitree_rl_gym g1_env.py（G1 策略观测定义） (Chinese)