Dead reckoning

Dead reckoning is a method of estimating where something is by starting from a known position and advancing it with measured or assumed motion (direction, speed, rotation or acceleration) over the time elapsed, without an outside position fix. In marine navigation, the U.S. American Practical Navigator defines it as determining a vessel's estimated position "by advancing from a known fix of position along the vessel's ordered course and speed". Because every small error in the measured motion is carried forward, a dead-reckoned position drifts away from the truth over time and has to be corrected periodically by an independent reference.

In virtual reality and augmented reality the term is used in two main senses. In tracking, it describes updating a headset's or controller's pose by integrating inertial measurement unit (IMU) readings between slower optical corrections; Oculus VR engineers described gyroscope integration on the Oculus Rift Development Kit as "a simple dead reckoning filter", and a 2017 study of the HTC Vive found that its pose is updated "primarily by inertial measurement units in the headset through dead reckoning". In networked simulations, multiplayer games and shared virtual environments, it describes extrapolating the motion of remote entities between network updates so that latency and bandwidth limits are less visible.

Reviewed 15 October 2026. Checked every claim against its cited source: Bowditch ch. 9, Woodman 2007, LaValle et al. 2014 and Niehorster et al. 2017 full texts, the OpenXR xrLocateSpace page, UploadVR, Game Developer, Meta, IEEE and arXiv pages, and the Foxlin, Harle, NPSNET and Singhal-Cheriton abstracts and Crossref metadata. About review dates.

Definition

In navigation, dead reckoning (abbreviated DR) gives an approximate position. The American Practical Navigator, published by the U.S. National Geospatial-Intelligence Agency, notes that it can be used both to determine where a vessel is now and to project where it will be, and that it does not allow for errors such as helmsman error, compass error, current and wind. Navigators compare the dead-reckoned position with a known fix to work out the forces acting on the ship, and the manual recommends keeping dead reckoning as a cross-check of GPS/GNSS positions shown on electronic chart systems when paper charts are the backup. It states that the method's most important use is projecting the ship's position into the immediate future to avoid hazards.

The same structure applies in electronic tracking. A system starts from an initial state and repeatedly adds the motion it measures, so errors in each measurement accumulate. Oculus VR researchers called this accumulated dead-reckoning error "drift error" and noted that sensing beyond the gyroscope is needed to reduce it (see Drift (tracking)). Practical systems combine dead reckoning with an absolute reference, such as a position fix at sea, gravity and the magnetic field for orientation, or cameras and base stations for position, a combination usually handled by sensor fusion.

Inertial dead reckoning

An inertial navigation system (INS) tracks orientation by integrating rate-gyroscope signals once, and tracks position by rotating accelerometer signals into a global frame, subtracting gravity and integrating twice. Oliver Woodman's 2007 Cambridge technical report on strapdown systems built from MEMS sensors shows why this drifts quickly. A constant accelerometer bias, when double integrated, produces a position error that grows with the square of time, and accelerometer white noise produces a random walk in position whose standard deviation grows with time to the power 3/2.

The larger problem is orientation. A small tilt error in the computed orientation projects part of the acceleration due to gravity onto the horizontal axes, and that residual is then integrated twice. In Woodman's tests of a simple INS running on a stationary Xsens Mtx IMU, the average position drift after 60 seconds was 152.67 m, of which only 1.76 m was vertical; he attributes nearly all of it to gyroscope noise propagating through the orientation estimate. Using magnetometers and Xsens's own sensor fusion algorithm for orientation cut the average error after 60 seconds to around 5 m. The report concluded that sub-meter accuracy for more than a minute was not yet possible with a MEMS-based INS.

Domain constraints are the other main correction tool. Woodman cites InterSense's NavShoe, which assumes a walking person's foot has zero velocity while it is on the ground and uses these moments as zero-velocity updates to reset velocity drift; he also notes the weakness of such assumptions, since NavShoe would fail if the wearer rode an escalator.

Reported drift figures

System	Conditions	Reported position error	Source
Simple strapdown INS on an Xsens Mtx IMU	Stationary, 60 s, INS attitude algorithm	152.67 m average drift
Same IMU with magnetometer-aided orientation (Xsens fusion)	Stationary, 60 s	About 5 m average drift
Oculus VR Rift Development Kit IMU, accelerometer double integration	One minute of running time	Exceeds 500 m
Same tracker with a seated two-link torso and head kinematic model	One minute of running time	Within 1.1 m
InterSense NavShoe (foot-mounted IMU)	Outdoor use	Meter-level error

Pedestrian dead reckoning

Pedestrian dead reckoning (PDR) applies the idea to people on foot, mainly for indoor tracking. Robert Harle's 2013 survey in IEEE Communications Surveys & Tutorials describes PDR systems as using inertial and other sensors, often combined with knowledge about walking, to track a user's movements. It develops a taxonomy of PDR systems and reviews techniques for step detection, step characterization, inertial navigation and step-and-heading dead reckoning. It also covers particle filters that incorporate building maps, hybrid systems that use absolute position fixes, and smartphones as PDR sensors. Harle concludes that PDR alone can give good short- to medium-term tracking under certain conditions but needs regular absolute fixes for long-term use.

A widely cited foot-mounted system is InterSense's NavShoe, described by Eric Foxlin in a 2005 paper in IEEE Computer Graphics and Applications. The sensor could run all day from a small battery and be tucked into the shoelaces. Foxlin wrote that outdoors a user would barely notice NavShoe's meter-level error, but that indoors computer vision would still be needed for precise tracking of a head-mounted display; the approximate position from the shoe could reduce the database search space for the vision system, and the foot unit could also act as a portable azimuth reference for a handheld or head-mounted tracker. A 2020 review by Johan Wahlström and Isaac Skog, written fifteen years after Foxlin's paper, credited it with popularizing the zero-velocity update and stated that how to design a robust zero-velocity detector was still an open research question.

Use in VR and AR tracking

Head orientation

Inertial dead reckoning was central to head tracking on the Oculus Rift Development Kit. In their ICRA 2014 paper, Steven M. LaValle, Anna Yershova, Max Katsev and Michael Antonov of Oculus VR described the Rift Development Kit's sensor board (an Invensense MPU-6000 gyroscope and accelerometer and a Honeywell HMC5983 magnetometer, reporting at 1000 Hz). They integrated each gyroscope reading as a quaternion rotation, which they called "a simple dead reckoning filter", and named the accumulated difference from the true orientation "drift error". Calibration could not remove every gyroscope error term, so they corrected tilt drift by comparing the measured direction of gravity from the accelerometer with the estimate, and corrected yaw drift with the magnetometer, using small gains so that corrections stayed imperceptible to the user. Earlier, Eric Foxlin's 1996 paper for the IEEE Virtual Reality Annual International Symposium had addressed the same trade-off for inertial head trackers: gyroscopes avoid jitter, latency, interference and line-of-sight problems but drift slowly, so he fused them with drift-free inclinometers in a Kalman filter.

Head tracking also uses forward extrapolation. The same Oculus paper described predictive tracking that projects the current orientation forward over the latency interval, assuming either a constant angular rate or a constant angular acceleration. For a 20 ms prediction interval with a user turning the head at up to about 240 degrees per second, the average error was 1.46302 degrees with no prediction, 0.19395 degrees with constant rate and 0.07596 degrees with constant acceleration. The authors presented these predictive methods as a way to reduce effective latency, the lag between a head movement and the corresponding change on the display (see Motion-to-photon latency).

Position

Position is much harder to dead reckon than orientation. LaValle and colleagues noted that a fixed accelerometer offset gives quadratic error growth after double integration, and that constant linear velocity is unobservable to an IMU of this kind. With plain double integration their position error exceeded 500 m after one minute; constraining the head to a seated torso and neck model kept it within 1.1 m, although the authors wrote that this method was only effective over a few seconds and was insufficient on its own. They wrote that their latest prototypes instead combined the inertial methods with a camera observing infrared LEDs on the headset. Later positional tracking systems kept this hybrid pattern. UploadVR's 2019 explainer of VR tracking states that dead reckoning is "how every VR headset and controller tracks itself" from moment to moment, that integrating accelerometer data twice magnifies and accumulates even the smallest error, and that optical systems such as Oculus Constellation or camera-based SLAM exist to correct this drift by providing a reference. In camera-based inside-out tracking, the article explains, the headset's rotation and acceleration from the IMU are compared with how static features in the room appear to move (see Visual-inertial odometry).

For the original HTC Vive, Diederick Niehorster, Li Li and Markus Lappe reported that the tracked pose is updated mainly by the headset's IMUs through "dead reckoning (path integration)" because this allows much higher update rates, while the Lighthouse base stations limit and correct the build-up of error at 120 Hz. When both base stations were out of view, the Vive did not continue on inertial data alone: the screen turned gray and the reported position and orientation stopped changing.

Tracking loss

Dead reckoning is also used to bridge short gaps in optical tracking. The OpenXR specification says that during tracking loss runtimes should continue to provide inferred or last-known poses, which "can, for example, be based on neck model updates, inertial dead reckoning, or a last-known position"; the runtime keeps the position-valid flag set but may clear the position-tracked flag to show that the value is estimated. At Oculus Connect 5 in 2018, Oculus developer relations engineer Gabor Szauer told developers porting apps to Oculus Quest that if controller tracking is lost for only "a second or two" they can usually "fake it pretty easily"; UploadVR suggested this could mean simple prediction algorithms. In 2019 Ready at Dawn said that for Echo VR on the inside-out tracked Oculus Rift S it could use "the in-game context, knowledge about our movement system, and the internal IMU sensors to accurately predict what the player is doing", addressing moments when a controller is outside the view of every headset camera. UploadVR interpreted this as using the accelerometer to detect when a player pushes off a surface with an untracked hand. A 2025 Meta developer blog post said that Double Jack, the studio behind the hand-tracked conducting game Maestro, built dead-reckoning features that estimate baton movement when the player's hands disappear from tracking.

AR navigation

Handheld AR navigation systems have used pedestrian dead reckoning where GPS is unavailable. A 2017 survey of AR navigation by Gaurav Bhorkar describes the indoor system of Alessandro Mulloni, Hartmut Seichter and Dieter Schmalstieg, in which the user scans way-points printed on the floor and the phone estimates position in between by counting steps; because this dead reckoning loses accuracy over time, the user rescans a way-point along the route. The survey also describes outdoor smartphone AR navigation that falls back on dead reckoning from orientation and accelerometer data when GPS is temporarily unavailable.

Networked simulation and multiplayer

In networking, dead reckoning hides latency and saves bandwidth by letting every participant extrapolate the motion of entities it does not control. Jesse Aronson's 1997 Gamasutra article describes the technique as implemented in the Distributed Interactive Simulation (DIS) protocol, which grew out of DARPA's SIMNET project. When an entity is created, its owner broadcasts an entity state protocol data unit (PDU) with its kinematic state (including position, velocity and acceleration) and an identifier naming the dead reckoning algorithm the other nodes should use. Each node extrapolates the entity locally, while the owner runs the same algorithm and sends a new PDU only when the extrapolated and true states differ by more than an agreed threshold, and at least once every five seconds. Aronson described the trade-offs: large thresholds cause visible jerkiness when corrections arrive, small thresholds increase traffic, DIS simulations typically add smoothing to ease the corrections, and dead reckoning offers little gain if all entities behave unpredictably all the time. The DIS application protocols are standardized as IEEE 1278.1; the 2012 revision defines the PDUs exchanged among simulation applications, grouped into protocol families such as entity information/interaction. The IEEE lists that revision as inactive-reserved since 30 March 2023.

Early networked VR research built on this work. NPSNET-IV, from the Naval Postgraduate School, was described in 1994 as the first 3-D virtual environment to combine the IEEE 1278 DIS application protocol with IP multicast for multiplayer simulation over the Internet. In 1995 Sandeep Singhal and David Cheriton argued that both sending position at the local frame rate and dead-reckoning protocols that use higher-derivative information require considerable bandwidth and at times behave poorly, and proposed a position-history protocol in which update packets carry only position and remote hosts extrapolate from several past updates to render remote players smoothly. Their paper framed the problem in terms of distributed virtual reality systems, which must render animated entities controlled on other machines in real time to support interaction between users across the network.

See also

• Drift (tracking)
• Sensor fusion
• Inertial measurement unit
• Kalman filter
• Complementary filter
• Predictive tracking
• Visual-inertial odometry
• Positional tracking
• Head tracking
• Latency