What an inertial sensor measures
An inertial sensor reports how a body is moving, without needing to look at anything outside it. The two main kinds used in hobby robots are the accelerometer and the gyroscope. Small chips that combine both, often called inertial measurement units or IMUs, are cheap and common. Some also include a magnetometer, which senses magnetic field direction like a compass.
Understanding inertial sensors helps with self-balancing robots, DIY drone builds, camera gimbals and any project that must know its orientation.
Accelerometers: sensing push and gravity
An accelerometer measures acceleration along one or more axes. The detail that surprises many people is that it also senses gravity. A sensor sitting still on a table does not read zero; one axis reads about one g, the pull of gravity, because the sensor is being pushed upward by the table.
That is useful. When the sensor tilts, the gravity reading is shared differently among its axes, and a bit of trigonometry converts those shares into a tilt angle. A stationary accelerometer is therefore a decent way to find which way is down.
Strengths
- Gives an absolute reference to gravity, so it does not wander over time.
Weaknesses
- It cannot tell gravity from movement. If the robot accelerates, brakes or bumps, those forces add to the reading and the apparent tilt is wrong.
- Motor vibration makes the signal noisy.
- The readings are jumpy over short intervals.
Gyroscopes: sensing rotation rate
A MEMS gyroscope measures how fast something is rotating, in degrees or radians per second, around each axis. It does not directly report an angle. To get an angle, the controller adds up the rotation rate over time, a process called integration. If the robot rotates at 10 degrees per second for 3 seconds, the angle changed by about 30 degrees.
Strengths
- Smooth and responsive over short periods.
- Not disturbed by linear acceleration in the way an accelerometer is.
Weaknesses
- Tiny constant errors, called bias or offset, build up as the readings are integrated. The estimated angle slowly wanders, even when the robot is perfectly still.
Understanding drift
Suppose a gyroscope reports a false rotation of just 0.1 degrees per second when still. Integrated over one minute, that false rate produces an apparent angle change of 0.1 × 60 = 6 degrees. Left running for ten minutes, the error would reach 60 degrees. The numbers are only illustrative, since real offsets vary by sensor and temperature, but the pattern is real: a small steady error grows without limit.
Drift is reduced in several ways:
- Calibrate at rest. Average the gyro readings while the robot is still and subtract that offset.
- Allow for temperature. Offsets can shift as the chip warms up, so some designs wait for warm-up or recalibrate.
- Correct with another source, such as an accelerometer or magnetometer, which is the main idea behind sensor fusion.
Combining the two: sensor fusion
The two sensors have opposite habits. The accelerometer is right on average but noisy in the short term. The gyro is smooth in the short term but wanders over the long term. Combining them uses each for what it does best.
A complementary filter is the simplest common method. Each cycle, the new angle is mostly the previous angle updated with the gyro’s rotation, with a small correction pulling it toward the accelerometer’s angle. Trusting the gyro in the short run and the accelerometer in the long run gives a stable, responsive estimate. A single tuning number sets how much weight each gets.
A Kalman filter is a more formal approach that estimates the angle and the gyro bias together, weighting each measurement by how noisy it is believed to be. It can perform better but needs more care to set up. Many hobby projects get acceptable results from a complementary filter, and some chips and libraries offer built-in fusion, though it is worth knowing what the output represents.
Heading, rotation about the vertical axis, is harder. Gravity gives no information about it, so an accelerometer cannot correct yaw drift. A magnetometer can help, though nearby motors, wires and steel often disturb it.
From tilt to balance
A self-balancing robot is like balancing a broom on your palm. When it leans forward, the wheels must drive forward under the lean to catch it. The loop is:
- Read the inertial sensor and estimate the tilt angle.
- Compare it with the upright target.
- Command the wheels to move in the direction of the lean, with strength depending on how large and how fast the tilt is.
- Repeat quickly, often hundreds of times per second.
The corrective step is usually a feedback controller, and PID control is the standard tool. The loop has to be fast and steady, because a late correction can make things worse. Wheel encoders are often added so the robot also knows its speed and position.
Practical build notes
- Mount rigidly. A sensor attached with soft tape picks up vibration. Use solid mounting hardware, and mount it near the centre of rotation where practical, with a known orientation.
- Reduce vibration. Balance wheels and motors where possible; some builders use soft isolation for the sensor board, though it has to be done carefully.
- Clean power. Noisy supplies degrade readings, so steady voltage regulators matter, and decoupling capacitors near the chip help. Many modules include their own, but poor motor wiring can still inject noise.
- Check the bus and voltage. Most inertial chips talk over a serial bus and have specific voltage limits. The datasheet states them, and reading datasheets carefully avoids damaged parts.
- Plan for scale settings. Chips offer selectable ranges: a narrow range gives finer resolution, a wide range avoids clipping during sudden knocks.
- Log data. Printing or recording raw readings while moving the robot by hand is a good debugging strategy for separating sensor problems from control problems.
Limits and expectations
Inertial sensing estimates orientation; it does not directly measure position. Integrating acceleration twice to get position drifts quickly and is not reliable for more than short moments. For position, add encoders, distance sensors or other references. Hobby inertial chips are inexpensive and good enough for balancing and tilt sensing, but they should not be expected to behave like expensive navigation equipment.
If a robot can injure someone when it falls or races away, include a dependable way to cut motor power, independent of the balancing code.