Are Today’s MEMS Gyros “Good Enough”?

Big improvements in precision may require new applications and market dynamics.


The gyroscope market is heating up, fueled by increasingly autonomous vehicles, robots, and industrial equipment, all of which are demanding greater precision and ever-smaller devices.

Gyroscopes historically have been a staple in navigation for years. However, classic designs are macro-mechanical, and high-performance units can be very expensive. For lower-performance applications, micro-electromechanical systems’ (MEMS) inertial measurement units (IMUs) have become increasingly popular due to their small size and lower cost. However, performance of these MEMS devices has been steadily increasing, allowing them to take on tougher roles. Most of them still need help from other sensors to cross-check their results, but they’re getting better.

“Academics have been for many, many years demonstrating all kinds of cool gyro architectures that could be used to create a low-drift gyro,” said Alissa Fitzgerald, founder and managing member at A. M. Fitzgerald and Associates. “But there haven’t been any commercial forces or market pull to bring that technology out of the academic world and into the real market.”

There are many tools available for navigation. IMUs, consisting of accelerometers and gyroscopes, are important, but less-expensive MEMS versions are assisted by global navigation satellite-system (GNSS) signals and other inputs from cameras, radar, and lidar – as well as magnetometers – to correct for drift. As performance improves, they need less correction. Still, it’s unlikely that they’ll be completely self-sufficient for critical applications anytime soon.

IMUs measure movement by tracking changes in both linear and rotational inertia. They consist of at least two different types of sensors, accelerometers and gyroscopes. These sensors can be built using either traditional macro-scale techniques or, increasingly, by using MEMS technology, where semiconductor techniques create sensors on a silicon chip.

Accelerometers provide a measurement of linear motion by sensing linear acceleration. That acceleration can be integrated to provide speed, and double integration provides position. A single accelerometer will measure changes along one axis. Combining three accelerometers with orthogonal axes can then measure 3D changes in the direction of motion.

Gyroscopes (or simply gyros) measure changes in orientation. This is distinct from the direction of travel, since one can change directions without changing which direction one is facing. Likewise, one can change orientation (effectively spinning) while moving in a straight line.

Fig 1: The left image shows linear acceleration to the left with no rotation — skidding. The middle image shows rotation with no linear acceleration — spinning. The right image shows both linear acceleration to the left and rotation – turning. Source: Bryon Moyer

Ideally, combining accelerometers or gyroscopes for all three axes on a single chip can provide better axis alignment, and combining accelerometers and gyroscopes on a single chip can ensure the best axis alignment between the two types of sensors. Magnetometers cannot be included in a monolithic integration because their construction isn’t compatible with accelerometers and gyroscopes.

While monolithic integration is possible, IMUs often are made with separate co-packaged chips for the different sensor types. “Most IMUs are assembled using three individual chips (a three-axis gyro, a three-axis accelerometer, and a three-axis magnetometer),” said Fitzgerald. “High-performance IMUs might contain three individual single-axis gyro chips and could contain five to nine chips.” In this case, calibration is required to adjust from any deviations from orthogonality.

Whether or not an IMU contains a single chip, or combines separate accelerometer and gyroscope chips, each axis for each sensor is counted. This provides a so-called six-axis sensor (if there is no magnetometer). IMUs with a three-axis magnetometer are considered to be nine-axis sensors.

Dead reckoning, drift, grades, and prices
Dead reckoning refers to the process of measuring one’s location by tracking the changes in position and orientation over time. Travel in a particular direction at a given speed will yield a new position. Further travel from that position will give a new position.

The challenge, however, is that no measurement is without error. Each measured position will have some error, and the measurement of the next position will also have error — and that error will compound the error of the first position. These errors can accumulate to the point where one’s actual location can deviate substantially from the calculated location. “It is typical to define the application grade of these systems by the performance, as measured by bias stability (or bias instability or drift),” according to Yole Développement’s Dimitrios Damianos, a technology and market analyst, and Guillaume Girardin, director of the photonics, sensing and display.

Drift is measured in degress/hr, or the number of degrees of orientation error accumulated in an hour.

Fig. 2: Error accumulation due to drift. Source: Bryon Moyer

Both accelerometers and gyroscopes have errors, but overall drift has been dominated by gyroscopes. This has resulted in the grading of IMUs according to the applications to which they’re suited. There are no official definitions of grades, only generally accepted ones, and they may vary. One version shown below goes by decades, but the middle range is subject to interpretation. “The three terms that get muddled a lot are industrial, navigation, and tactical,” said Fitzgerald. “Those are used interchangeably to refer to performance of around a half a degree per hour.”

Fig. 3: Bias stability by market grade. Source: Alissa Fitzgerald, from Position, Navigation, and Timing Technologies in the 21st Century: Integrated Satellite Navigation, Sensor Systems, and Civil Applications)

Yole Développment has published a list of grades, as well.

Fig. 4: IMU grades as defined by bias instability (or stability). Source: Yole Développement

While gyroscope drift dominates the definition of grade, Yole points out that other factors matter, as well. “Stability is not the only parameter that counts,” said Damianos and Girardin. “There are other specifications such as resistance to vibration and shock, bandwidth, broad operation temperature range, stability over temperature, size/weight/power, etc. You can’t use a gyro for navigating a ship and the same system for guiding a missile.”

The most demanding applications still require mechanical sensors and are extremely expensive. On the opposite end of the scale are inexpensive commercial applications where accuracy is less critical and the consequences of inaccuracy are less dire.

Fig. 5: IMU prices as defined by bias instability. Color indicates technology: RLG = ring-laser gyro; HRG = hemispheric resonator gyro; FOG = fiber-optic gyro. MEMS is gradually improving the grades it can achieve. Source: Yole Développement

Older mechanical models are giving way to newer technologies, including MEMS technologies that are gradually improving. MEMS units may encroach upon some of what FOG technology can address today, but they are unlikely to replace HRG and RLG technologies in the future. That said, FOG technology is projected to grow enough that, even while losing the lower end to MEMS, it will still see growth and hold its own.

Fig. 6: IMU technology evolution for high-end applications. The lightbulb indicates when the idea occurred; the dot indicates when it became available (giving the indicated time-to-market). Source: Yole Développement

MEMS applications and improvements
MEMS technology has lowered the price of lower-end IMUs, which are used primarily in commercial (like smartphones) and industrial applications. “Until recently, none of the big electronic companies was really interested in navigation grade gyros,” said Fitzgerald. “They were making gyros for smartphones and game controllers, which are, frankly, crappy gyros for vehicle-navigation purposes. We haven’t yet seen the new MEMS architectures get fully developed. And part of the reason for that is that it’s a tough pitch to have an investor give you $50 million to commercialize these exciting new technologies.”

But Yole noted that silicon MEMS gyroscopes have improved to the point where they can address navigation-grade applications. “Generally, all technologies are steadily improving towards more stability and better performance,” said the Yole team. “Some technologies, like MEMS, which were initially at the commercial performance level, have moved to industrial grade and recently to tactical.”

According to Bosch product managers Peter Spoden and Michael Rupp, MEMS gyroscopes have cut their drift in half, with further improvements expected. “Bias instability of 10°/h was industry standard for MEMS for many years,” they said. “Today, 1 to 5 °/h is feasible with MEMS. For the future, the range below 1°/h seems accessible.”

Fig. 7: Comparison of drift for silicon-based MEMS (“SiMEMS”) vs. other technologies. (Bias instability (stability) was calculated using Allan variance method at room temperaturure, with no acceleration or vibration. In-run bias stability, run-to-run bias stability is calculated as the quadratic sum of previous errors depending on mission profile. In-run is a single run, while run-to-run checks repeatability. Source: Yole Développement

So will MEMS technologies continue to move into the higher-end applications? With the advent of the autonomous vehicle, interest has been renewed in higher-quality sensors priced for the automotive market. But while one company appears to have made some strides in that direction, most makers believe that further improvement isn’t really critical due to the fact that IMUs are typically combined with other technologies to correct for errors.

Sensor fusion
While dead reckoning would not be possible with MEMS IMUs for any extended period, MEMS gyroscopes are still positioned for use in safety-critical applications like self-driving car navigation. That’s because gyroscopes (and IMUs in general) are but one navigation technique available.

The primary source of position information remains GNSS signals, like GPS. “Many use cases rely on GPS for drift correction with a Kalman filter or the like, so they don’t have to worry so much about long term drift,” said Jeff Miller, application engineering consultant at Mentor, a Siemens business. That fails when a vehicle enters a tunnel or a parking garage where the satellites delivering the signals are no longer “visible.” In that case, IMUs can provide short-term location information until the GNSS signal is re-acquired.

“If you’re going to be able to tolerate a five-second fuzz-out of GPS and you’re going 50 to 70 mph, you’re going to need a pretty good gyro,” said Fitzgerald. “If you are worrying only about a one-second drop-out, maybe you can get away with a less accurate gyro.” But the longest road tunnel in the Alps, for instance, is the Gotthard Road Tunnel, which is about 17 km (10.5 mi) long. At 60 mph, that would mean being away from the GNSS signals for slightly more than 10 minutes, and that’s a long time for a MEMS IMU to be on its own.

Magnetometers also are used to help correct for gyroscope drift. Magnetometers measure orientation, like gyroscopes, but they do so relative to magnetic north. That provides a relatively stable reference point, so algorithms can look at both gyroscope and magnetometer results to decide on the orientation.

Unlike GNSS signals, which trades off with an IMU when unavailable (meaning that either one or the other is in use), here the two sensors are constantly fused together to achieve a result that is, hopefully, more accurate than either sensor on its own. This is a specific example of what’s called sensor fusion. It merges the results of numerous sensor outputs to generate a better combined result.

That said, magnetometers also have their weaknesses – largely in their response to large features that can affect the magnetic field. Referred to as “magnetic anomalies,” metallic vehicles, elevators, and even large rock outcrops can distort the magnetometer’s reading. In applications where such anomalies persist – in proximity to large motors or within an electric room, for example – magnetometers cannot be reliably used.

Outside those applications, however, anomalies may be transient. So the gyroscopes, in addition to being cross-checked by the magnetometer, also help to cross-check the magnetometers. There is no one fixed algorithm for rejecting anomalies. “We use both heuristic and first-principles approaches to magnetic anomalies,” said Rohit Seth, founder and CTO at Micron Digital.

Beyond that, there are numerous other technologies that can help with location services. IMUs provide an “inside-out” view, where the moving sensors determine their own position. External signals, like GNSS, provide an “outside-in” view, where external factors tell the moving item where it is. Within a shopping mall, for instance, various WiFi or Bluetooth beacons can triangulate the position of a shopper and their smartphone. Together, the IMU in the smartphone and the beacon signals can provide positioning accurate enough to locate the shopper within a store.

This approach could also be used on the roadways. “We could start putting beacons in our roadways, in our infrastructure,” noted Fitzgerald. “You can put beacons in a tunnel.”

Of critical concern with autonomous vehicles is the fact that there’s little margin for drifting off course. “If you’re driving your car down the highway, you probably can’t afford to drift more than 10 or 20 centimeters before you’re putting yourself at risk of hitting an adjacent car,” said Fitzgerald. “And maybe 30 centimeters is the maximum.”

This is where advanced driver-assist systems (ADAS) are particularly important. Such systems equip a vehicle with numerous additional sensors — cameras, radar, and lidar, in particular. Combining these sensor signals using AI algorithms for recognizing items, both expected and unexpected, provides an enormous sensor-fusion solution.

“Environmental recognition is always necessary for self-driving cars,” said the Bosch team. “The IMU can never be used alone and must always be part of a complete system, consisting of other devices such as lidar, radar and software. As part of such a system, the IMUs of today are sufficient.”

They may, however, be less than sufficient for a “go-anywhere” vehicle that may not stay on city roads with well-painted lines. But autonomy in such far-flung locales is also more of an open question (especially if there is no cellular communication). “It would be very difficult to have an autonomous car going in terrain where you don’t have those lane markers or traffic lights or stop signs to recognize,” said Fitzgerald.

Which ones are appropriate for an automobile may not be obvious. “We need to separate the ADAS vehicles (with increasing level of autonomy) and robotic cars (or robotaxis), which are already fully autonomous but geographically limited,” said the Yole team. “Robotaxis do not have cost constraints for the moment, due to their different business model (Mobility as a Service). But they have high requirements on performance.”

With ADAS-equipped cars, the purchase price is the primary economic consideration. Therefore, pricing of the IMUs must be below the $100 level, according to Yole, giving them accuracy of a few centimeters of drift in 20 to 30 seconds at 60 km/hr. By contrast, robotaxis, may tolerate a higher price for the accuracy – as much as $10,000 for a system that drifts only a few centimeters per kilometer driven.

Lower-cost vehicles that need to navigate may still see higher-quality MEMS IMUs as too expensive. “When you’ve got a car that’s going to cost $50,000 to $100,000, spending $300 for a MEMS gyro doesn’t seem like a big deal,” said Fitzgerald. “It starts to fall apart for people who want to put these kinds of gyros into drones and robots and products that are not going to support a $300 price. They could pay for a $30 MEMS gyro, but $300 starts to be too much.”

Still, some of those smaller systems might be able to cope with less accuracy. “If a robot is going only three miles an hour, you can probably get away with a cheaper gyro because you’re not trying to navigate with so much precision,” she added.

One might even question whether, given all of the cross-checking sensors, an IMU is still required for some driving applications. “The progress made by the combination of others sensors (radar, cameras, lidar, etc.) raises the question of the usefulness of inertial systems in the long term,” the Yole team speculated. “At some point, all others sensors might be able to calibrate the positioning of the car without such sensors.”

Crowd-sourcing sensors
While sensor fusion typically combines the results of different sensors, there’s one more take on the notion from Micron Digital. The company combines the outputs of an array of three to six sensors using a simple neural network that, it claims, can bring drift close to zero. The idea of the array is that several identical sensors are measured, and the final decision comes out of that combination, not from any of the sensors in isolation. “The IMUs look at each other to estimate drift,” said Seth. “IMUs are correcting other IMUs.”

These arrays initially were built on a single module. “We started with commonly available sensors arranged in a macro-configuration,” said Seth. “This configuration is well tuned to third-party sensors. The macro system-on-module (SOM) can be condensed down to an approximately 19mm x 19mm footprint.”

The company also is working on smaller-footprint versions. “We are designing our own substrate version of the sensor so that it can be fabricated in under a 5-mm x 5-mm footprint,” he added. “There are intermediate designs that use an assembly of multiple substrates in a single SoC package.”

It’s not as simple, however, as a numeric aggregation like an average. The neural network must be hand-tuned. “We’re close to a general solution, but for now we need to tune per application,” said Seth. It’s also tuned for each brand of sensor, since they all work somewhat differently. They say this isn’t a fixed algorithm, but rather that it does dynamic calibration and decides on the fly which of the sensors to trust. There’s also an output stage that inhibits any unrealistic movements.

It’s still unclear how much further MEMS gyroscopes will improve. For the most part, they’re getting gradually better. But there is little pressure for them to improve for the biggest obvious application, automobiles, due to the massive amount of sensor fusion being designed into such cars.

If they were able to achieve yet higher grades of operation, then they might reduce costs at the high end, but many of those applications come with stringent requirements and lower volumes, making a return on that investment uncertain. Dramatic improvements will require a demanding new application.

Related Articles
Sensor Fusion Challenges In Cars
As more pieces of the autonomous vehicle puzzle come into view, the enormity of the challenge grows.
Understanding SLAM (Simultaneous Localization And Mapping)
How computer vision systems track movement.
Software-Defined Vehicles
The electrification of cars makes all sorts of things possible.
Pivoting Toward Safety-Critical Verification In Cars
Experts at the Table: Changing the automotive mindset; verification after manufacturing; security updates.
Who Owns A Car’s Chip Architecture
Carmakers and their suppliers compete for dominance, creating challenges across the electronics industry.


Darron Collins says:

Well written article, Bryon. This can be used to explain these technologies use cases for those attempting to understand the basics, before moving in to deeper individual topics. Thanks for your insight.

Leave a Reply

(Note: This name will be displayed publicly)