"The map is not the territory."
Alfred Korzybski, 1933
Most of what we call artificial intelligence meets the world at second hand. A language model reads text that a person already wrote; a vision model looks at pixels that a camera pipeline already demosaiced, white-balanced, and compressed. The world has been tidied up before the model ever sees it. Sensory AI starts one step earlier, at the device that measures reality directly and imperfectly: an accelerometer that feels a bearing wearing out, a radar chirp returning from a person behind a wall, a photodiode counting scattered photons, a strip of piezoresistive skin registering the first slip of a grasped cup. Before any of this becomes a clean array, it is a voltage, a count, a timestamp, and a physical process that does not care whether the model finds it convenient.
This book takes the position that perceiving the physical world through inertial, vibration, radar, lidar, depth, thermal, RF, tactile, and biomedical sensing is now a first-class AI discipline, not a preprocessing footnote on the way to a "real" model. These signals share a grammar that pixels and tokens do not teach: sampling rates and aliasing, the difference between a spectrogram and a raw waveform, calibration that drifts with temperature, and the blunt fact that a sensor can simply fail. A gyroscope has a bias that walks over time. A depth camera invents geometry at reflective edges. An ECG electrode loses contact and hands you a flat line that is not a flat heart. Our core stance throughout is to treat every sensor as an instrument with noise, bias, drift, and failure modes, and to model those explicitly, because a decision built on \( \hat{x} = x + \varepsilon \) is only trustworthy once you have taken \( \varepsilon \) seriously.
The timing is deliberate. Between 2023 and 2026 the field crossed a threshold that vision and language crossed years earlier: the arrival of sensor and time-series foundation models. Large models pretrained on inertial streams, physiological recordings, and heterogeneous industrial telemetry now transfer across tasks and devices with the same few-shot ease that made language models useful, and they do it while respecting the temporal and multi-rate structure of real measurements. That shift is part of why a book like this can be written now rather than a decade ago. It also raises the stakes: a foundation model fed a miscalibrated stream will generalize the miscalibration with great confidence.
So we welcome you to a single connected build, followed from end to end rather than in disconnected fragments. You will start at the raw sample, where a physical quantity becomes a number, and move through calibration, denoising, feature and representation learning, fusion across modalities, model design, uncertainty, and finally a decision that ships and is monitored in the field. Every chapter is hands-on, and every step keeps the whole pipeline in view, because in the physical world the last mile, a trustworthy deployed decision, is only as sound as the first sample it was built on. Turn the page, and let us start at the sensor.