Smartgloves finger data is noisy as hell — anyone found a sane cleanup workflow?

Pose-dependent variance is actually fixable if you're willing to do a short calibration pass. Capture noise samples per finger at three reference poses: full extension, mid-curl, and fully curled. Store three separate R matrices and interpolate between them at runtime based on the current curl estimate coming out of the filter.

More setup upfront, but the filter stops being overconfident at extremes, which is exactly where it matters most. Full extension and tight grip both tend to be high-noise states for most sensors, and a single flat R just can't handle both ends. The trickiest part is getting consistent reference poses across different hand sizes, but even rough three-point interpolation is way better than a flat R.

The wrist twist example gets at something: the real tuning knob in a 1D Kalman filter isn't the measurement noise covariance (R), it's the process noise covariance (Q). R is roughly fixed by your sensor spec. Q is where your assumptions about joint motion dynamics actually live.

For signals with high-variance bursts like pronation or fast finger extension, a fixed Q is exactly the wrong model. A simple adaptive variant works much better: track your recent innovation residuals (predicted vs. actual measurement), and when they spike, temporarily increase Q. The filter widens its uncertainty estimate so the update step can chase the movement. When residuals settle back down, Q drops and smoothing resumes. It's maybe 20 extra lines and it handles the "sudden movement should track, idle drift should smooth" requirement in a principled way rather than just picking a compromise fixed value and hoping.

EMA with a small alpha is usually my first pass too, but for finger curl specifically I've had better results with a simple one-dimensional Kalman filter. The EMA assumes noise is stationary, but finger tracking noise tends to vary with glove pressure and hand orientation. Kalman adapts its gain dynamically, tightening when the signal is stable and loosening during real movement. It's not much more code and the feel at slow controlled movements is noticeably better. EMA over-smooths exactly the subtle finger positions where precision actually matters.

the three-pose calibration is solid but don't assume symmetry between fingers. my index and pinky had noticeably different noise profiles at the same pose. the pinky was about 40% higher variance at full curl, probably just sensor placement variation inside the glove. sample each finger independently rather than averaging across the hand or you'll end up under-filtering the noisy ones.

also: the interpolation method between reference poses matters more than the number of poses. linear interp between three points still had visible jumps at the transitions. cubic spline over the same three samples was a lot smoother.

The Q/R split is where most Kalman filter tutorials lose people — they present both as noise parameters when in practice they're tuning completely different things. R is relatively straightforward to estimate from captured idle frames. Q is a modeling assumption about how much the true underlying motion can change between timesteps, which is much harder to pin down without just experimenting.

Something that helped me: running the filter offline over a few representative takes with different Q values and plotting output against the raw signal. Makes the lag/noise tradeoff far more visible than trying to tune it by feel during a live session. If you have archived takes lying around it's worth doing before you touch the production pipeline.

The Q vs R distinction is also where the physical intuition breaks down, because R is at least empirically measurable: hold your sensors still, record the variance, done. Q is asking "how much can the true state change between samples," which requires you to reason about the physics of the system you're tracking, not just the data.

For finger curl the awkward thing is that a single Q has to model both slow deliberate curls and fast snaps, and those have very different expected velocities. A Q loose enough to track a fast trigger pull is more permissive than you want for held poses, and a Q tight enough for held poses systematically lags on fast transients. There's no single value that works for both. That's usually where people end up looking at adaptive Q, or just accepting that fast movements will lag a bit.

The stationary noise assumption is what kills EMA for wrist twist data too. Rapid pronation/supination has way higher variance than idle drift, and a fixed alpha just can't adapt to that. The 1D Kalman approach is worth the extra setup time.

One thing I'd add: tune your process noise (Q) and measurement noise (R) separately per finger segment rather than using uniform values across the whole hand. Proximal joints move more aggressively and need a higher Q; distal tips tend to be noisier and benefit from a higher R. I treated them as uniform for way too long and got mediocre results everywhere instead of good results anywhere.

The hold still and measure R approach is solid, but finger sensors have pose-dependent noise. Variance at full extension isn't the same as when the finger is curled, and metacarpal joints behave differently from the proximal ones. Measuring in a neutral position and treating it as universal is a workable approximation, but for Smartgloves data I ended up measuring R in three representative poses (rest, open, fist) and averaging them. Still fully empirical, still beats guessing, just more representative of what you'll actually see during capture.