Sensors & feedback

A measurement is not the entire state

An encoder reports rotational position or position changes at a particular shaft. If it measures the motor side of a gearbox, converting to output angle requires the ratio and an offset; an output-side encoder observes the joint more directly. Some actuators have both. A reported motor angle alone cannot reveal every flexing bracket, loose fastener or foot contact.

A gyroscope measures angular rate. An accelerometer measures specific force, which is influenced by both gravity-related support and motion. A magnetometer senses the local magnetic field. Sensor fusion combines measurements and a model to estimate orientation. An IMU is not automatically a complete position tracker: integrating tiny acceleration errors can produce growing position error.

Choose axes before interpreting signs

Tape a small three-axis diagram to a book standing in for the robot torso: x forward, y left, z up. Now imagine the sensor is mounted sideways inside the book. The sensor’s x axis and the body’s x axis no longer describe the same direction. A sensor-to-body rotation is needed before the controller can interpret the measurements correctly.

Roll, pitch and yaw can describe orientation, but their meaning depends on rotation order and frame convention. A quaternion uses four values to represent rotation. Libraries disagree about whether the scalar comes first or last, so [w,x,y,z] and [x,y,z,w] are not interchangeable arrays. Record both the order and which frame the rotation maps from/to. You do not need to derive quaternion algebra to recognize that contract.

Projected gravity is a useful balance cue

The audited walking controller uses the gravity direction expressed in the body frame. For a level body whose z axis points up, a normalized downward gravity vector is [0,0,−1]. As the body tilts, components appear on the other axes. This tells the controller about tilt without requiring a worldwide heading reference.

That vector is not the raw accelerometer reading. A stationary supported accelerometer commonly measures an upward specific-force response, while the modeled gravity vector points downward. The exact sensor output mode matters. Learn to identify whether a field means raw acceleration, gravity, linear acceleration with gravity removed, or an orientation estimate before connecting it to a policy input.

Freshness is part of correctness

A beautifully formatted packet can still be wrong for the current moment. Keep the time of measurement, time of receipt and age at use separate when possible. A controller running fifty decisions per second has a 20 ms decision period, but the acceptable sensor age is an engineering choice that must be tested, not automatically equal to one period.

The early BDX-R paper and current deployment code use different sensor generations. The audited later implementation expects an MTi-3-oriented serial path; the earlier BNO055 setup is not a drop-in substitution. The audit also found missing freshness checks. These observations become requirements for the later mock-hardware lab rather than excuses to guess a sensor board or orientation.

The age of a reading is the time it is used minus the time it was sampled. If sampling happened at 5.00 seconds and use happens at 5.08 seconds on the same clock, the reading is 0.08 seconds old. There are 1,000 milliseconds in one second, so that age is 80 milliseconds. Compare timestamps only after confirming they share a clock or have a known conversion.