humanoidsdata.com

Search

Search datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Hardware & control

Proprioception

Proprioception is sensing of a robot’s own internal configuration and motion rather than the external scene. For a humanoid it commonly includes joint positions and velocities, actuator effort or torque, and inertial measurements of body rotation and acceleration. These signals support state estimation and feedback control but do not, by themselves, directly describe nearby objects or terrain.

Also known as: Robot proprioception, Proprioceptive sensing

Updated

What counts as proprioceptive sensing

The Springer textbook Elements of Robotics separates proprioceptive sensors, which measure the robot itself, from exteroceptive sensors, which measure the surrounding environment. Encoders, motor-state feedback and inertial measurement units are common proprioceptive sources; cameras and range sensors are exteroceptive.

Force and contact signals sit near the boundary because they reveal an external interaction through onboard hardware. Robotics papers do not always classify them consistently, so a dataset should list the actual channels rather than rely on the word “proprioception” alone.

Why humanoids need state estimation

Joint encoders describe the articulated shape but cannot alone determine a floating humanoid’s motion through the world. Inertial measurements and kinematic contact assumptions help estimate base orientation, velocity and pose. An IROS humanoid state-estimation study used joint encoders and multiple body-mounted IMUs because the common assumption of a flat, stationary support foot breaks during heel-to-toe motion.

Learned locomotion also depends on these signals. The ANYmal sim-to-real study used proprioceptive observations and an actuator model to transfer dynamic policies from simulation to hardware.

What training data must declare

Proprioceptive streams need timestamps, units, sign conventions, sensor frames, calibration and latency. Records should distinguish raw readings from filtered or model-estimated states, and commands from measured responses.

Simulation adds another trap: exact base velocity, contact state or disturbance forces may be available to the trainer but not directly measurable on the robot. Marking these privileged signals prevents a deployable observation set from being confused with information used only during training.

Sources