A walking robot knows where its joints are and what its accelerometers read. This proprioceptive data — no cameras, no lidar, no GPS — is enough to estimate the robot's position and orientation in principle. In practice, the standard approach learns a measurement model from data, assumes Gaussian noise, and fuses everything through a Kalman filter.
The problem is the Gaussian assumption. Real proprioceptive noise isn't Gaussian. Foot slippage, terrain compliance, and learned-model errors produce noise distributions with heavy tails and asymmetries that Gaussian filters can't represent. When the filter assumes Gaussian noise and encounters non-Gaussian reality, it either becomes overconfident (underestimates uncertainty) or drifts (accumulates bias it can't correct).
The alternative: don't assume any distribution at all. Instead, characterize the measurement noise as a set-coverage statement — a guarantee that the true state lies within a computable set with some probability, without specifying the shape of the noise distribution. This is weaker than a full probabilistic model but more honest. It says “the answer is somewhere in here” rather than “the answer has this probability density.”
Integrating set-coverage measurements with a Gaussian filter produces a hybrid: the filter propagates a Gaussian state estimate forward in time (which is computationally cheap), but corrects it using set-membership constraints (which are distribution-free). The result maintains consistency — the estimated uncertainty actually contains the true state — while avoiding the drift that comes from distributional misspecification.
The conceptual lesson: in sensing, an honest statement of ignorance outperforms a precise statement of the wrong thing. The Gaussian model provides more information, but the information is wrong. The set model provides less information, but the information is true.
Note (2026-11-06): I checked this essay against the raw arXiv abstract of Kulkarni, Poulakakis and Huang, 'Proprioceptive-only State Estimation for Legged Robots with Set-Coverage Measurements of Learned Dynamics' (2603.18308, v1). The abstract says the Gaussian assumption on learned measurement models 'can easily break down with limited training data', rendering estimates inconsistent and potentially divergent; the authors characterize the noise with distribution-free set-coverage statements used with a Gaussian filter, and validate in simulation and on two real-world quadrupedal datasets, where the method stays consistent and is not prone to drift. My specific noise sources (foot slippage, terrain compliance), my account of how the set constraints correct the Gaussian state, and the general conclusion that an honest statement of ignorance outperforms a precise statement of the wrong thing are my own. I checked the abstract and author list only.