RoughSense: Lightweight Terrain-Induced Rover Vibration Prediction Using Point Clouds and IMU Feedback

Scientific Impact Assessment: RoughSense

1. Core Contribution

RoughSense proposes a lightweight pipeline for predicting terrain-induced rover vibration by fusing exteroceptive (LiDAR point cloud) and proprioceptive (IMU) sensing. The core idea is a two-stage system: (1) a geometric vibration proxy computed as the RMSE of RANSAC plane-fitting residuals on local point-cloud patches, and (2) an online correction of this geometric prediction via Recursive Least Squares (RLS) using the actually-measured IMU vibration during traversal. The stated problem — that exteroceptive descriptors predict *before* contact but are indirect proxies, while proprioceptive signals measure *after* contact — is well framed, and the RLS bridge is a sensible, computationally cheap way to reconcile the two. The explicit target is space/subterranean robotics under tight compute, power, and communication budgets.

The contribution is essentially an engineering integration of well-established components (RANSAC plane fitting from Fischler & Bolles 1981; RMSE roughness proxies from Krüsi et al. and Hamner et al.; RLS from classical adaptive filtering). None of the individual pieces are new; the novelty lies in the specific combination and the online adaptation framing for a resource-constrained deployment. The authors are appropriately candid that their aim is to *correct* rather than replace geometric analysis.

2. Methodological Rigor

The design is reasonable but modest. Strengths: the paper includes a sensible frequency-domain preprocessing step (stationary vs. moving experiments to identify and notch-filter rover self-induced vibrations at ~27.5 Hz), speed normalization of the IMU vibration magnitude, and a genuine multi-environment evaluation (lunar analogue, lake shore, underground mine). The prediction-error analysis evaluates predictions at cells *s* grid-cells ahead and compares against subsequently-measured IMU scores, which is a meaningful predictive evaluation rather than a trivial in-place fit. Statistical significance is assessed with paired Wilcoxon signed-rank tests across distance bands, and computation-time profiling on a Jetson Orin (single CPU core) substantiates the "lightweight" claim (234 µs for correction).

Weaknesses are notable. The evaluation is essentially self-consistency: the "ground truth" is the IMU-derived score, and the corrected prediction is fitted to track exactly that signal — so improved agreement is partly built into the objective rather than validated against an independent physical vibration reference. There is no comparison against the LSTM baseline (Waibel et al.) that the related work explicitly cites as the alternative, nor against any learned method — so the central claim of "competitive at far lower cost" is asserted rather than demonstrated. Calibration relies on manually chosen min–max thresholds from "barely traversable" areas, introducing subjectivity. The linear RLS mapping is acknowledged to miss nonlinear terrain–robot dynamics. Only a single robot platform (Leo Rover) is tested.

3. Potential Impact

The practical relevance is real: vibration-aware traversability is a genuine need for planetary and subterranean rovers, and a method that runs in ~44 ms on a single CPU core without GPU is attractive for power-limited platforms. However, the impact is likely incremental. The method is a useful, deployable module for a niche community (space/field robotics traversability) rather than a conceptual advance. It is unlikely to redirect research directions or become a widely-cited primitive. A public GitHub repository is provided, which aids adoption.

4. Timeliness & Relevance

The topic is timely — the DARPA SubT Challenge, lunar lava tube exploration, and the "new space" movement all motivate lightweight autonomy. Online adaptation of geometric priors with proprioceptive feedback is an active theme. The paper addresses a real bottleneck (compute/power constraints preventing deployment of heavy learned models), which is a genuine and current concern. (Note the arXiv ID and dates suggest a future-dated venue — ISPARO2026 — which does not affect the technical assessment.)

5. Strengths & Limitations

Strengths: Clear problem framing; genuine three-environment field validation, including a hard underground setting; honest and thorough limitations section; strong attention to real-time compute budget with actual profiling; sensible signal-processing hygiene (self-vibration filtering, speed normalization); reproducibility aided by released code.

Limitations: No baseline comparison against competing methods (learned or geometric-only variants beyond the ablation of point-cloud-only vs. corrected); evaluation is somewhat circular (fitting to and evaluating against the same IMU proxy); single platform; manual calibration thresholds; linear model limits; degraded performance at long range in the mine (the most operationally relevant regime for the stated use case); modest quantitative gains in the hardest environment (25% in the mine, and no significant long-range improvement there).

Additional Observations

Overall, RoughSense is a competent, well-executed field-robotics engineering paper with honest evaluation and real deployability appeal, but limited conceptual novelty, absent external baselines, and niche scope constrain its expected scientific influence.