A sensor model, also called an observation or measurement model, is a probabilistic description of how a robot’s sensor readings relate to the underlying state of the world, expressing the likelihood of an observation given a hypothesised state. It captures sensor characteristics such as noise, resolution, range limits, and failure modes, allowing a robot to weight evidence appropriately when fusing measurements. Sensor models are central to Bayesian state estimation, where they form the update step that corrects predictions using incoming data.
Overview
- In recursive state estimation the sensor model supplies the measurement-update (correction) step that refines a predicted belief.
- It complements the Motion Model, which supplies the prediction step.
- Well-calibrated sensor models let a robot weight conflicting evidence during Sensor Fusion.
Mechanisms
- Characterise the sensor’s noise distribution and systematic biases empirically.
- Express the conditional likelihood of readings given state, including beam, feature, or pixel-level models.
- Incorporate the likelihood into a Kalman Filter or Particle Filter update.
- Validate against ground truth and recalibrate as hardware drifts.
Applications
- Robot Localisation against a known map.
- Simultaneous mapping and localisation in SLAM.
- Multi-sensor Sensor Fusion and Robot Perception.
- Building consistent maps via Mapping.