Exteroception

• Also called: exteroceptive sensing, external sensing
• Category: Sensing › Exteroceptive
• Tags: sensing, perception, sensor-classification, proprioception, perceptive-locomotion

Sensing of the environment outside the robot, with cameras, lidar, radar, sonar and similar sensors, as opposed to proprioception, which senses the robot's own joints and motion.

Definition

Exteroception is a robot's sensing of the world outside its own body: the shape, appearance and position of the ground, obstacles, objects and people around it. The term comes from physiology, where Charles Sherrington distinguished exteroceptive senses such as vision and hearing from proprioceptive senses of the body's own position and movement. In robotics the pair is a standard way to classify sensors. Exteroceptive sensors include cameras, depth cameras, lidar, radar, ultrasonic rangefinders, microphones and GNSS receivers; proprioceptive sensors include joint encoders, motor current sensing and IMUs.

The boundary is not always sharp. An IMU senses the robot's own motion but also the direction of gravity, and touch and force sensors measure contact with the environment yet are usually grouped with proprioception in legged locomotion work. A practical test is whether the signal mainly describes the robot's internal state or the world around it. Exteroceptive data is typically high-dimensional, arrives with some latency, and is affected by lighting, weather, dust, reflections and occlusion. Proprioceptive data is low-dimensional, fast and dependable, but it reveals nothing about the environment until the robot touches it.

Much of robot design is about combining the two. In legged locomotion, blind controllers that use only proprioception are robust but react to terrain only after contact, while perceptive controllers add exteroceptive inputs, such as height samples from an elevation map, so that the robot can place its feet ahead of time. Because exteroception can fail or mislead, for example on tall grass, snow, glass or water, robust systems estimate how far to trust it and fall back on proprioception when the two disagree. Vehicle datasets such as TartanDrive record exteroceptive streams, such as front camera images and top-down height and color maps, alongside proprioceptive signals so that models can learn how terrain affects the vehicle's motion.

Why it matters

Exteroception is how a robot learns about anything it has not yet touched: obstacles, the terrain ahead, objects to grasp and people nearby. Knowing what each exteroceptive sensor can see, where it fails and how to combine it with proprioception is a basic part of designing any robot that works in an unstructured environment.

Key applications

• Obstacle detection and avoidance for mobile robots
• Perceptive legged locomotion on stairs and rough terrain
• Object detection and pose estimation for manipulation
• Mapping, localization and SLAM
• Terrain classification for off-road driving
• Detecting and tracking people for safety

In practice

A quadruped team adds a perceptive layer to a walking controller that already works blind. Depth cameras and a lidar feed an elevation map around the robot, and the policy receives height samples around each foot in addition to joint states and IMU data. In testing, the robot now climbs stairs it previously tripped on, but it stumbles in tall grass, where the map shows a solid surface that is not there. The team retrains the policy with randomized and corrupted height maps so that it learns to rely on proprioception when foot contacts disagree with the map, and they log both streams to find terrain where the exteroceptive input misleads the robot.

Common misconceptions

• Exteroception means cameras — it covers any sensing of the outside world, including lidar, radar, sonar, microphones and GNSS
• More exteroceptive sensors always make a robot more robust — each adds failure modes and latency, and the controller needs a way to tell when the external view is wrong
• Exteroception and proprioception are competing choices — most capable robots combine them, using proprioception for fast, dependable feedback and exteroception for anticipation
• Every sensor fits cleanly into one category — IMUs, contact and force sensors sit near the boundary, and how they are classified depends on what the signal is used for

Who should know this

• Perception and sensing engineers
• Legged and mobile robot locomotion developers
• Sensor fusion and state estimation engineers
• Robotics students learning how sensors are classified

Related technologies

• Lidar
• Stereo and Time-of-Flight Depth Cameras
• Radar
• Ultrasonic Rangefinders
• Elevation Mapping
• Multimodal Sensor Fusion

Related terms

• Proprioception
• LIDAR
• Depth Camera
• RGB Camera
• Elevation Map
• Sensor Fusion

Related projects

• Open3D: Modern library for 3D point cloud processing, mesh operations, and 3D machine learning
• PCL: Industry-standard C++ library for point cloud processing in robotics perception
• OpenMind OM1: AI-native operating system for robots with modular perception, cognition, and action
• Patchwork++: Fast and robust ground segmentation for 3D LiDAR point clouds from KAIST
• Vision Pilot: Autoware Foundation's open-source L2 ADAS stack from a single camera
• RAI: Vendor-agnostic agent framework for ROS 2 robots and physical AI