CtrlK
See also: Robotics, SLAM, Autonomous driving, Sensor fusion, Augmented reality, Lidar
An inertial measurement unit (IMU) is an electronic device that measures and reports the specific force, angular rate, and (in many configurations) the orientation of a body relative to a local reference frame. It does this by combining the readings of one or more accelerometers and gyroscopes, sometimes augmented with magnetometers, barometric pressure sensors, and temperature sensors. IMUs are the central inertial sensing element behind nearly every modern motion-aware system, from smartphones and drones to humanoid robots, virtual reality headsets, autonomous vehicles, guided missiles, and submarines.[1][2]
In its purest form an IMU is a sensor package that produces raw inertial data: linear acceleration along three orthogonal axes from accelerometers, and angular velocity around three orthogonal axes from gyroscopes. Modern integrated IMUs frequently bundle additional functionality, including built-in calibration, temperature compensation, on-chip sensor fusion using a Digital Motion Processor (DMP), and digital interfaces such as I2C, SPI, and UART. When orientation is computed on board from the raw sensor stream, the resulting system is usually called an Attitude and Heading Reference System (AHRS).[3][4]
The transition from precision mechanical and optical inertial sensors costing hundreds of thousands of dollars to MEMS-based silicon chips priced under one US dollar has reshaped robotics, mobile computing, and consumer electronics. Each smartphone now carries an IMU, every drone uses one for stabilization, every modern AR/VR headset depends on one for low-latency head tracking, and every advanced humanoid uses several for balance and motion estimation. The maturation of the IMU is therefore one of the foundational hardware enablers of contemporary Robotics and embodied AI.[5][6]
An IMU is, at minimum, a tightly co-located sensor cluster that provides the inertial measurements needed to integrate motion over time. The number of sensing axes determines the descriptor that is commonly attached to the device.
The phrase "degrees of freedom" (DoF) used in the IMU industry refers to the count of independent sensor channels in the package, not to the geometric degrees of freedom of a rigid body. The two dominant configurations are:
The raw signals from an IMU are deceptively simple but require careful interpretation:
| Sensor | Quantity measured | Typical units | Reference frame |
|---|---|---|---|
| Accelerometer | Specific force (gravity plus inertial acceleration) | m/s² or g (1 g ≈ 9.81 m/s²) | Body |
| Gyroscope | Angular velocity | rad/s or deg/s (dps) | Body |
| Magnetometer | Magnetic flux density | µT or gauss | Body, mapped to Earth field |
| Barometer | Atmospheric pressure | hPa | Local atmosphere |
| Temperature | Internal die temperature | °C | Sensor die |
A stationary IMU placed flat on a desk reports an accelerometer reading near (0, 0, 9.81) m/s² because it is being pushed up by the desk against gravity, and a gyroscope reading near (0, 0, 0) rad/s. The accelerometer measures specific force rather than acceleration, which is the source of long-running confusion among new IMU users.
Three adjacent terms are often confused. The distinctions matter because they determine how much of the navigation problem is solved on the chip versus by the host system.
| Term | Outputs | Typical algorithms | Examples |
|---|---|---|---|
| IMU | Raw accel, gyro (and mag, baro) | None onboard, host computes orientation | InvenSense MPU-6050, Bosch BMI088 |
| AHRS | Roll, pitch, yaw (orientation) | Onboard fusion, e.g. Kalman or complementary | Bosch BNO055, Honeywell HG2800 |
| INS | Position, velocity, attitude | Strapdown integration, often GNSS-aided | Honeywell HG5700, Northrop Grumman LN-200 |
An AHRS extends an IMU by computing orientation continuously, while an inertial navigation system (INS) integrates the IMU output to produce a full position-velocity-attitude (PVA) state. INS designs are categorized as strapdown when the IMU is rigidly fixed to the vehicle, with all rotation handled in software, or gimbaled when the IMU is mechanically isolated on a stable platform. Modern systems are almost universally strapdown because of cost, reliability, and computational tractability.[9]
Micro-electromechanical systems (MEMS) sensors dominate the consumer and commercial IMU market. A MEMS accelerometer typically uses a small proof mass suspended on silicon springs whose displacement under acceleration is measured capacitively. A MEMS gyroscope is most often a vibrating-structure (Coriolis) device in which a tuning-fork or ring proof mass is driven into resonance, and Coriolis-induced motion perpendicular to the drive axis is read out capacitively or piezoelectrically.
Three manufacturing strategies came to dominate MEMS IMUs and accounted for roughly 94% of all units integrated into smartphones in the 2019–2022 generation: STMicroelectronics, TDK InvenSense, and Bosch Sensortec. STMicroelectronics held about 50% of the smartphone IMU market, TDK InvenSense about 27%, and Bosch Sensortec about 17%. The three players differ in process choices: TDK InvenSense moved to a 90 nm node with copper metal layers and Al-Ge eutectic wafer bonding to combine the ASIC and MEMS dies into a single silicon stack, while STMicroelectronics and Bosch Sensortec keep the MEMS and ASIC as separate dies in an LGA package.[5]
The consumer IMU market was projected to reach about US$838 million in 2026, with roughly 5% compound annual growth between 2020 and 2026.[5]
For higher accuracy classes, optical gyroscopes replace vibrating MEMS structures with light-based interferometry:
Three-axis magnetometers in IMUs are typically Hall-effect or anisotropic magnetoresistive (AMR) devices. Asahi Kasei's AK8963 magnetometer was paired with the InvenSense MPU-6500 to form the historically influential MPU-9250 nine-axis MARG sensor.[8] Barometric pressure sensors based on piezoresistive or capacitive MEMS membranes are added in 10-DoF IMUs for altitude estimation, particularly in drones.
IMUs are graded by accuracy, primarily characterized by gyroscope in-run bias stability (the lowest noise floor of the bias as it slowly wanders) and angle random walk (the high-frequency noise that accumulates as a random walk when integrated to angle). The classes from lowest to highest accuracy are consumer, industrial, tactical, navigation, and strategic.[1][13]
| Grade | Gyro bias stability | Position drift (unaided) | Cost | Typical use |
|---|---|---|---|---|
| Consumer | 10 to 100 deg/hr | Many km/min | < $1 to $10 | Smartphones, fitness trackers, toys |
| Industrial | 1 to 10 deg/hr | Hundreds of m/min | $50 to $500 | Robotics, drones, AR/VR |
| Tactical | 0.1 to 1 deg/hr | About 1 nautical mile/hr | $1,000 to $50,000 | Guided munitions, tactical UAVs, autonomous trucks |
| Navigation | 0.001 to 0.1 deg/hr | Less than 1 nm/hr | $50,000 to $500,000 | Commercial aviation IRUs, naval ships, long-range cruise missiles |
| Strategic | < 0.001 deg/hr | Meters per hour | $1M+ | Submarines, ICBMs, deep space probes |
The distinction between grades is not arbitrary marketing. A gyro that drifts at 100 deg/hr is unusable for long-baseline navigation but adequate for screen rotation, whereas a strategic-grade unit drifting under 0.001 deg/hr can sustain dead-reckoning navigation across an oceanic submarine patrol that may last weeks.[9]
IMU manufacturers and integrators characterize the noise statistics of accelerometers and gyroscopes using the Allan variance, originally developed by David W. Allan to measure the frequency stability of precision oscillators. The IEEE Standard 952-2020 (and predecessor 647-2006) defines the use of Allan variance for ring laser gyros and inertial sensors more broadly. Plotting the Allan deviation versus averaging time on a log-log graph reveals signature slopes that correspond to identifiable noise sources:[13][14]
For consumer IMUs, the bias instability is typically reached after seconds to a few minutes of averaging. For tactical FOG IMUs, hundreds of hours of static data are often required to fully characterize bias instability and slow drifts.[13]
A bare IMU produces noisy raw data. Useful orientation, velocity, and position estimates require fusing those raw measurements (often together with other sensors) through filters that exploit the complementary error characteristics of each modality. Gyroscopes are accurate over short time scales but drift at low frequencies, while accelerometers are noisy over short time scales but provide a stable absolute reference (gravity) at low frequencies. Magnetometers add an absolute heading reference at low frequencies but are highly susceptible to local magnetic disturbances.[15]
| Algorithm | Year | Computational cost | Strengths | Weaknesses |
|---|---|---|---|---|
| Complementary filter | 1960s (general), 1990s (IMU) | Very low | Trivial to implement, no matrix math | Single tuning gain, no covariance |
| Mahony filter | 2005, 2008 | Low | Explicit complementary form, P+I gyro bias correction | Two gains to tune |
| Madgwick filter | 2010 | Low | Quaternion gradient descent, single tuning parameter, on-line magnetometer compensation | Less accurate than EKF in dynamic conditions |
| Extended Kalman filter (EKF) | 1960s | Moderate | Probabilistic, fuses arbitrary sensors, propagates covariance | Linearization errors near singularities |
| Unscented Kalman filter (UKF) | 1997 | Higher | No Jacobians, better for non-linear models | More tuning, higher CPU |
| Multi-state constraint Kalman filter (MSCKF) | 2007 | High | Tightly coupled visual-inertial, low memory | Complex implementation |
| Factor graph (iSAM, GTSAM) | 2008 onward | High but incremental | Sparse optimization, supports loop closure | Latency, batch-style processing |
| Learned bias / hybrid neural | 2022 onward | High (training), low at inference | Captures non-Gaussian errors, generalizes across temperatures | Requires labeled data |
Empirical comparisons consistently find that the Madgwick and Mahony filters yield orientation errors only about 0.4 degrees worse than a well-tuned Kalman filter in benign conditions, while running an order of magnitude faster and using no matrix operations. A typical reported overall mean error across these filters is about 3.4 degrees.[15][16]
The Madgwick filter, introduced by Sebastian Madgwick in his 2010 internal report and the 2011 IEEE ICORR paper "Estimation of IMU and MARG orientation using a gradient descent algorithm," represents orientation as a quaternion to avoid Euler-angle singularities and uses a gradient-descent step on the accelerometer (and optionally magnetometer) error to correct gyro integration. Its single tunable parameter beta encodes the gyroscope measurement error rate. The filter compensates magnetometer distortion online and applies a gyro bias drift correction.[16][17]
Madgwick's filter dominated open-source AHRS implementations during the 2010s, in part because Madgwick released both the algorithm and reference C and MATLAB source code under permissive licenses. The Reefwing-AHRS, Adafruit AHRS, and many ROS packages bundle either Madgwick or Mahony out of the box.
The Mahony filter (Robert Mahony, Tarek Hamel, Jean-Michel Pflimlin, 2008) uses an explicit complementary filter formulation with proportional and integral feedback in the body frame. The proportional gain (k_P) corrects fast attitude error from the accelerometer/magnetometer reference, while the integral gain (k_I) corrects slow gyro bias drift. Mahony is among the fastest IMU filters in absolute compute time, making it attractive for microcontrollers running at hundreds of hertz.[15]
The extended Kalman filter and its variants (UKF, EKF on manifolds, error-state EKF, invariant EKF) underpin most production IMU fusion in robotics and autonomous vehicles. The error-state Kalman filter is particularly common in INS, where it estimates a small-state correction (errors in position, velocity, attitude, and IMU biases) around a high-rate strapdown integration of the raw IMU. This decoupling allows the inertial integration to run at thousands of hertz on a microcontroller while the EKF correction runs at the rate of aiding measurements such as GNSS or vision.[9][18]
The single most influential advance for low-cost IMUs since the late 2000s has been their fusion with cameras to form visual-inertial odometry (VIO) and visual-inertial SLAM (SLAM augmented with inertial measurements). The IMU provides high-rate motion estimates between camera frames and resolves the absolute scale that monocular cameras cannot measure, while the camera provides drift-free position references through visual landmarks. Mourikis and Roumeliotis's 2007 multi-state constraint Kalman filter (MSCKF), adopted in Apple's ARKit and many autonomous vehicle stacks, was a key step toward consumer-grade VIO.[19] Modern systems include OKVIS, VINS-Mono, ROVIO, ORB-SLAM3, and Kimera, all of which depend critically on a calibrated IMU.[20]
Fusion can also extend to lidar (lidar-inertial odometry such as LIO-SAM, FAST-LIO), wheel encoders (wheel-inertial odometry on ground robots), GNSS (loosely or tightly coupled GPS/INS), and event cameras. The 2025 trend has been to learn IMU bias models with deep neural networks (LSTM and Transformer architectures, and most recently diffusion models) to keep VIO accurate during long visual outages, with research demonstrating that learned bias prediction can sustain pose estimation when the camera is occluded for extended periods.[21][22]
The foundational work on tightly coupled IMU/GPS integration for autonomous land vehicles was published by Salah Sukkarieh, Eduardo Nebot, and Hugh Durrant-Whyte in IEEE Transactions on Robotics and Automation in 1999. They demonstrated a high-integrity navigation loop combining low-cost strapdown IMUs with standard or carrier-phase GPS, including fault detection for low-frequency IMU bias and high-frequency GPS multipath errors. This work informed nearly all later commercial GPS/INS products used in autonomous vehicles, surveying robots, and precision agriculture.[23]
IMU calibration covers intrinsic parameters (axis scale factors, axis misalignments, biases, temperature dependence) and extrinsic parameters (the rigid transform between the IMU and other sensors). The Kalibr toolbox developed at ETH Zürich by Paul Furgale, Jörn Rehder, Jakob Maye, Janosch Nikolic, Roland Siegwart, and colleagues became the de facto open-source standard for camera-IMU calibration, with extensions to LIDAR-IMU and multi-IMU calibration. Kalibr formulates a continuous-time joint calibration using B-splines for the time-varying IMU states and embeds the temporal offset between sensors as an optimization variable.[24]
In 2010 InvenSense (acquired by TDK in 2017) introduced the Digital Motion Processor (DMP) inside the MPU-6050 and later MPU-9150 and MPU-9250 chips. The DMP is a small embedded processor that runs sensor fusion, gesture recognition, and step counting at typically 200 Hz directly on the IMU silicon. Its primary purpose is to offload timing-sensitive integration from the host CPU and to provide quaternion or Euler-angle outputs through the same I2C bus used to fetch raw data. The DMP was a major contributor to the rise of cheap nine-axis fusion in hobbyist robotics and consumer wearables.[8]
Later Bosch Sensortec products such as the BMI270 followed a similar pattern, integrating gesture, context, and activity recognition with an integrated step counter at sub-milliamp current draw, suitable for always-on wearable applications.[25]
In higher-end devices, IMU fusion has moved off the IMU silicon and onto a dedicated low-power sensor hub processor (e.g., the Apple Motion coprocessor M-series, the Qualcomm Sensor Hub, the Apple R1 chip in the Apple Vision Pro, and the NVIDIA Drive sensor stack in autonomous vehicles). These chips run sophisticated multi-sensor fusion across IMU, camera, lidar, and barometer streams while keeping the application processor in deep sleep, and they typically meet hard real-time deadlines for AR and AR/VR head tracking.
The IMU industry spans roughly six orders of magnitude in price and accuracy. The same article and wiki may discuss devices that cost less than US$1 alongside units priced above US$1 million. The principal manufacturers per grade are summarized below.
| Manufacturer | Headquarters | Notable IMU products | Grade focus |
|---|---|---|---|
| STMicroelectronics | Switzerland | LSM6DSO, LSM9DS1, LIS302DL (original iPhone) | Consumer, industrial |
| TDK InvenSense | USA / Japan | MPU-6050, MPU-9250, ICM-20948, ICM-42688-P | Consumer, industrial |
| Bosch Sensortec | Germany | BMI088, BMI270, BMI323, BNO055 | Consumer, industrial |
| Analog Devices | USA | ADIS16470, ADIS16500, ADIS16577 | Industrial, tactical |
| Epson | Japan | G320, G354, G355, G370, G570 | Industrial, tactical |
| Honeywell | USA | HG1700, HG1900, HG2800, HG5700, GG1320 RLG | Tactical, navigation |
| Northrop Grumman | USA | LN-100, LN-200, LN-251 | Tactical, navigation |
| Safran (formerly Sagem) | France | Sigma 95N, Geonyx | Navigation |
| KVH Industries | USA | 1750/1775 IMU | Tactical |
| Inertial Labs | USA | IMU-P, INS-D, INS-DM | Industrial, tactical |
| iXblue | France | Octans, Phins | Navigation (HRG) |
The STMicroelectronics LIS302DL was the first IMU-class accelerometer used in a smartphone of global impact. Apple shipped it in the original iPhone in 2007, where it enabled automatic screen orientation as the user rotated the device, a feature widely credited with shaping public expectations of touch devices.[26][27]
The InvenSense MPU-6050 (released in 2011) was a milestone in low-cost six-axis MEMS IMUs, with onboard DMP, a unit price under US$5, and an enormous community ecosystem in Arduino, Raspberry Pi, and ROS. The follow-on MPU-9250 added an Asahi Kasei AK8963 magnetometer in a single QFN package. After both parts were end-of-lifed by TDK, the ICM-20948 became the recommended replacement nine-axis IMU.[8]
The Bosch BMI088 is a workhorse high-vibration six-axis IMU widely deployed in drones such as DJI quadcopters and humanoid robots, with mechanical design specifically aimed at suppressing structural vibration. The BMI270 targets always-on wearable applications, drawing about 685 microamps in full operation while providing on-chip activity classification and a step counter. Bosch's BNO055 is a 9-axis IMU + Cortex-M0 fusion engine that exposes Euler angles directly over I2C, popular with hobbyists who want to skip filter implementation entirely.[25][28]
Analog Devices' ADIS family (ADIS16470, ADIS16500, ADIS16575, ADIS16577, etc.) targets industrial and tactical use with calibrated, temperature-compensated IMUs and SPI interfaces. The ADIS16500 specifies an 8.1 deg/hr in-run gyro bias stability and 125 μm/s² accel bias stability, sufficient for short-baseline GPS/INS in field robotics. The 2026 launch of the ADIS16577 at about US$450 disrupted tactical pricing and forced competitors to refresh their product lines.[29]
Epson sensing's G320N, G354, G355, G370, and G570 quartz MEMS IMUs are used in survey-grade GNSS/INS receivers, antenna stabilization, and rail/road inspection.[30]
Northrop Grumman's LN-200 is the canonical tactical-grade FOG IMU, with three solid-state fiber-optic gyros, three silicon MEMS accelerometers, and the lowest gyro/accelerometer white noise and highest MTBF in its class. It is widely used on tactical UAVs, missile guidance, and as the inertial heart of high-end GPS/INS such as the NovAtel SPAN-LN200.[11]
Honeywell's HG2800 family combines next-generation MEMS accelerometers with closed-loop FOGs for electro-optical gimbal stabilization, AUV/UAV navigation, and silent acoustic-sensitive operation. Honeywell's GG1320 ring laser gyroscope, with more than 500,000 units delivered, is the foundation of inertial reference units in nearly every commercial airliner.[10][12]
Navigation-grade FOG and HRG products such as Honeywell's HG5700, Northrop Grumman's LN-251, and iXblue Phins serve commercial aviation inertial reference units (IRUs), naval surface ships, submarines, and long-range cruise missiles.
Every modern smartphone ships with a 6-DoF or 9-DoF IMU. Beyond the original screen-rotation use case in the 2007 iPhone, IMUs now power image stabilization (working with the camera ISP to compensate for hand shake), pedometry, fall detection, augmented reality (Apple ARKit, Google ARCore), gaming, and indoor positioning. Apple's M-series motion coprocessors and the Qualcomm Sensor Hub keep the IMU streaming continuously while the application processor sleeps, enabling always-on activity tracking.[26]
Fitness wearables (Apple Watch, Garmin, Fitbit, Whoop) use IMUs for step counting, exercise recognition, heart-rate motion artifact rejection, sleep staging, and crash/fall detection. Apple Watch fall detection in particular uses IMU-derived acceleration and rotational signatures to trigger an automatic emergency call after a hard fall.
Multi-rotor and fixed-wing drones rely on IMUs for the inner attitude control loop, typically running at 1 to 4 kHz. Open autopilots such as PX4 and ArduPilot use IMUs of varying grades: hobbyist quadcopters use a single Bosch BMI088 or InvenSense ICM-42688-P; commercial inspection drones may carry triple-redundant ADIS16470-class units; military UAVs may carry a Honeywell HG1700 or Northrop Grumman LN-200.
The latency budget for Augmented Reality and VR head tracking is brutal: humans perceive lag of about 20 ms or less as breaking immersion. IMUs are the only sensors fast enough to drive head pose at the required hundreds of hertz, with cameras and depth sensors providing slower drift correction. The Apple Vision Pro ships with four IMUs, twelve cameras, and a dedicated R1 sensor processor that delivers head tracking and 3D environment mapping at low enough latency to pass through the user's hand without visible lag. Spatial audio is also keyed to IMU-derived head pose so that audio sources stay locked in space as the user looks around.[31][32]
The Meta Quest line, Microsoft HoloLens 2, and standalone VR headsets all follow the same pattern: high-rate IMU integration plus camera-based VIO drift correction. Inside-out tracking with no external base stations is now the consumer-VR standard, and it is built on this IMU + camera fusion stack.
In Robotics, IMUs are a near-universal standard sensor. The Robot Operating System (ROS) defines sensor_msgs/Imu, a message type that carries a timestamp, an orientation quaternion (with covariance), an angular velocity vector (with covariance), and a linear acceleration vector (with covariance), all in standardized SI units (rad/s and m/s²). When the orientation cannot be computed by the driver, the convention is to set the first element of the orientation covariance matrix to -1, signaling consumers to ignore the orientation field. This message type has become the lingua franca for IMU data exchange between drivers, fusion nodes, and visualization tools across thousands of ROS packages.[33]
Humanoid robots (Boston Dynamics Atlas, Tesla Optimus, Figure 02, Agility Digit, Unitree H1) use one or more IMUs in the torso and pelvis to estimate the pose of the floating base, the velocity of the center of mass, and the zero-moment point in the support polygon. Combined with joint encoders, contact sensors, and forward kinematics, these inertial estimates close the loop on whole-body balance controllers and walking gait generators. Quadrupeds (Boston Dynamics Spot, ANYmal, Unitree Go2) follow the same pattern.
Nearly every Autonomous driving stack uses an IMU as one node in a multi-sensor fusion graph that also includes wheel encoders, GNSS, lidar, radar, and cameras. The IMU provides the high-rate motion model that propagates the vehicle state between slower aiding measurements, which is essential during GNSS dropouts in tunnels, urban canyons, and parking structures. Tactical-grade IMUs such as the Honeywell HG1700 or Analog Devices ADIS16500 are common in robotaxi-class vehicles, while industrial-grade MEMS IMUs are used in driver assistance (ADAS) packages. Mining, agriculture, and long-haul trucking applications often use higher-grade FOG IMUs to maintain accurate localization over hours of GNSS-denied operation.
Guided munitions (GPS-guided JDAMs, cruise missiles, anti-ship missiles), hypersonic vehicles, ballistic missiles, fighter aircraft, transport aircraft, ships, and submarines all depend on inertial navigation built on tactical-, navigation-, or strategic-grade IMUs. In GPS-denied environments caused by jamming, spoofing, or natural blockage, INS ensures continuous navigation by relying solely on internal sensors. Strapdown INS architectures based on RLGs, FOGs, and HRGs are now the norm; gimbaled platforms persist mostly in legacy fleet upgrades.[9][34]
The principal challenges in IMU-based estimation are well understood and motivate most of the software stack on top of the raw sensor:
Key standards documents that govern IMU specification and testing include:
IMU manufacturer datasheets typically report Allan variance plots, bias instability and angle/velocity random walk in IEEE-standard units, plus temperature coefficients, vibration sensitivity (g-sensitivity for gyros), bandwidth, output data rate, and noise spectral density.