How can proprioceptive sensing characterize granular terrain during high-speed hopping?

Depth-only terrain models break down at impact

Most proprioceptive terrain sensing assumes granular resistance scales mainly with how deep the foot has penetrated. That assumption holds when a leg sinks slowly, but a hopping leg meets the ground at speed, and the paper's controlled hops show the resulting force–depth curve deviating systematically from a linear depth-only fit. The deviation is not random scatter; it clusters at specific moments of the stride.

The experiments used a vertically constrained, direct-drive hopper with an 80 mm circular foot, hopping on a bed of 300 µm glass beads prepared with a fluidized-bed technique before each trial. Five nominal touchdown speeds (0.22, 0.51, 0.83, 1.01 and 1.22 m/s) were crossed with compression stiffnesses of 2.50, 3.75 and 5.00 N/cm, with five trials per configuration and 60 trials total. Only the first stride of each trial was analyzed, so prior contacts would not compact the surface. A foot-mounted load cell supplied ground-truth force while motor current, encoders, IMUs and a time-of-flight sensor supplied the proprioceptive signals.

Two transients carry most of the error

In a representative stride, the foot did not stop instantly at contact as it would on rigid ground; granular yielding decelerated it over roughly 50 ms. The foot IMU recorded a sharp upward acceleration transient at impact, and the load-cell force rose to an early peak that was lower and broader than a rigid-ground impact. A second pronounced acceleration transient appeared at the compression–extension transition, when the controller raised virtual leg stiffness.

These two intervals — shortly after touchdown and after the stiffness switch — are exactly where a depth-and-speed force model left large residuals. The paper treats this co-location as the diagnostic clue: both intervals coincide with rapid foot deceleration, pointing to a force contribution tied to acceleration rather than to depth or speed.

Entrained grains act as an added mass

The paper separates terrain reaction force into a depth-dependent resistance, a velocity-dependent quadratic drag term, and an acceleration-dependent added-mass term. The added-mass idea is that a high-speed foot drags a conical region of grains beneath it, so accelerating the foot also accelerates that grain mass. Under constant-speed intrusion the acceleration term vanishes, which is why constant-speed tests alone cannot expose it.

To test this, the authors ran constant-speed intrusions from 0.022 to 1.1 m/s in 0.02 m/s increments and fitted a depth–speed force surface. Adding velocity dependence improved agreement but left the two transient residuals intact. Estimating an added-mass profile and multiplying it by measured foot acceleration reproduced the timing and shape of those residuals, though magnitudes differed modestly — a mismatch the authors attribute to uncertainty in gradient integration and bulk-model fitting. This is model-based evidence for added mass as a primary contributor, not a direct measurement of entrained grains.

A momentum observer removes inertial bias

Proprioceptive force estimated from motor current alone must carry the leg's own rigid-body inertia and gravity, so the raw quasi-static estimate mismatched the load cell. The paper adds a momentum observer that compensates for those terms, producing a corrected contact-force estimate from onboard sensing.

The effect on terrain inference was measured as error against the linear-actuator ground-truth granular resistance, k_GT. Across the touchdown-speed sweep at fixed compression stiffness and across the stiffness sweep at roughly 1.0 m/s touchdown speed, momentum-observer-corrected estimates sat closer to k_GT than the quasi-static estimator, reducing trial-level RMSE from 4.84 to 3.47 N/cm. Residual deviations remained, especially at high touchdown speeds where accelerations are large — the gap the next step targets.

Down-weighting high-acceleration samples closes the gap

Because the leftover error correlated with large foot acceleration, the regression was changed rather than the force model. Each sample receives an inverse-variance weight, and a sigmoid raises the assigned variance once the absolute foot acceleration exceeds a threshold a0. High-acceleration samples are treated as noisier measurements but are kept in the fit, so the regression leans on the calmer parts of the stride without discarding the transients.

With this acceleration-aware weighted least squares, the estimated resistance tracked k_GT more closely across both the touchdown-speed sweep and the stiffness sweep, and trial-level RMSE fell further from 3.47 to 2.35 N/cm. The two corrections do different jobs: the momentum observer fixes the force signal, while the weighting fixes how that signal is turned into a resistance value. The paper reports both as necessary for consistent estimates across locomotion conditions.

Where the result holds and where it stops

The evidence comes from one platform and one medium: a single-DoF vertical hopper with a flat circular foot on dry, uniform glass beads, with only the first stride of each trial analyzed. The authors explicitly frame extension to other leg morphologies, multi-degree-of-freedom contacts, and cohesive or fluid-saturated substrates as future work, noting that geometry- and direction-dependent force representations may then be required. The reported RMSE values are trial-level errors against a linear-actuator ground truth, not field performance on natural terrain.

Within that scope, the paper's practical priority is clear: when touchdown speed or a stiffness switch produces large foot accelerations, acceleration-dependent granular effects must be modeled explicitly. The comparison that matters is the ordering of the three treatments — quasi-static at 4.84 N/cm, momentum-observer-corrected at 3.47 N/cm, and momentum-observer plus acceleration-aware weighting at 2.35 N/cm. A reader deciding whether to add complexity to a terrain estimator should look first at whether their gait generates the acceleration transients this paper identifies; if it does, depth-based regression is the wrong baseline.

Sources used in this answer