Clearance-Confidence Potential Fields: Letting LiDAR Uncertainty Set the Safety Margin

A potential field controller steers a robot by treating the goal as a pit and obstacles as hills, then following the downhill slope at the robot's current position. It is cheap and reactive. It also trusts its sensors completely. When a LiDAR (laser rangefinder) misses returns on a dark box, or reports a phantom point from dust, the slope jumps, and the robot either swerves at nothing or drives into something it never saw. This proposal changes what the repulsive hill is built from. Instead of raw points, it uses a Gaussian process (GP), a Bayesian regression model that returns both a best guess and an error bar at every location, fitted to the current scan. The robot is then repelled by a pessimistic estimate of clearance: the estimated distance to the nearest obstacle minus a multiple of its uncertainty. Where the sensor has been reliable, margins shrink and the robot can pass through centimeter-scale gaps. Where returns are missing or contradictory, margins grow automatically.

GPs are usually considered too slow for this: exact inference costs $O(n^3)$ in the number of points. A potential field controller, however, only needs the posterior and its gradient at a handful of points on the robot's footprint, not over a map. That removes the prediction bottleneck that Fast Gaussian Process Occupancy Maps spends its effort pruning. What remains is a small local conditioning step once per scan. The empirical claim to test is narrow. The benefit should come from the uncertainty term, not from smoothing. It should therefore appear when sensor corruption is spatially uneven, as it is on real hardware, and vanish when noise is uniform, where a well-tuned fixed inflation radius should do just as well.

Core Claim

In cluttered ground navigation with spatially uneven sensor corruption (dropouts on specific surfaces, isolated phantom returns), we claim a repulsive potential defined on a GP lower-confidence clearance bound beats the same controller built on raw or filtered points with a tuned fixed safety margin. It should yield higher success and fewer deadlocks in tight passages, at control-loop rates on a CPU. The mechanism is that missing data raises GP variance locally. Per-beam filtering cannot represent this, because it treats an absent return as free space.

Several pieces of prior work shape this claim. Two lines motivate it. CARE attaches a classical repulsive field to vision policies using monocular depth, and its authors report false maneuvers when depth fails on reflective or dark surfaces, because the field is deterministic. The BARN challenge reports (2022, 2024; Autonomous Ground Navigation in Highly Constrained Spaces: Lessons Learned From the Third BARN Challenge at ICRA 2024 [Competitions]) show that physical success in centimeter-clearance mazes went to classical hierarchical planners or to learned policies wrapped in explicit geometric safety checks.

The closest methods are the following. COGIS fits a GP implicit surface to contacts and feeds its uncertainty into a sampling-based controller, but it removes spurious points with a slow derivative-free search and targets manipulation episodes. We need neither: free-space ray samples in the GP already down-weight phantom points, and the controller is a gradient law rather than a sampler. EGO-Planner shows repulsive gradients can be computed without a global distance field, but its obstacle anchors are deterministic.

The method reuses one established fact from Derivative Observations in Gaussian Process Models of Dynamic Systems and Bayesian Optimization with Gradients: the derivative of a GP is itself a GP. This gives closed-form spatial gradients of both the mean and the standard deviation.

One design is ruled out by construction. GPnn re-selects nearest neighbors for every query, which makes the surface discontinuous as the query moves. Taking gradients of such a surface produces force spikes. We avoid per-query neighbor switching and test that choice as an ablation.

Research Program

Each incoming 2D scan is converted into training data in the robot frame. Hit points get target value zero. Points sampled along each ray before its hit get positive targets equal to their distance to the hit, truncated at a sensing radius $\rho_0$. This gives the GP a signed-distance-like field $f(x)$ with posterior mean $\mu(x)$ and standard deviation $\sigma(x)$. The kernel lengthscale and noise level are fitted once offline on simulated scans and then frozen. We accept the hyperparameter rigidity that also limits OHSVGP and test whether it matters.

For scalability, points within a fixed radius of the robot at scan time are subsampled to a budget of $m$, and the Cholesky factor is computed once per scan. Between scans the conditioning set is fixed, so the field the controller sees is infinitely differentiable in position. Changes in the force across scans are driven only by new data, and we measure them directly.

The repulsive potential is placed on the lower confidence bound on clearance, $c(x) = \mu(x) - \kappa\,\sigma(x)$, where $\kappa \ge 0$ sets how pessimistic the robot is:

$$U_{\text{rep}}(x) = \tfrac{\eta}{2}\Big(\tfrac{1}{c(x)} - \tfrac{1}{\rho_0}\Big)^2 \ \text{for } c(x) < \rho_0, \quad 0 \text{ otherwise},$$

Here $\eta$ is a gain. The force is $-\nabla U_{\text{rep}}$, and $\nabla c = \nabla\mu - \kappa\nabla\sigma$ is available in closed form from the derivative GP. The control law is evaluated at a few footprint points. Setting $\kappa = 0$ recovers a smoothed-but-certainty-blind field; this is the key ablation.

We do not claim to solve local minima. The attractive term pulls toward subgoals on a Dijkstra global path, following the hierarchical structure that Pokle et al. and the BARN 2022 winner found necessary. The analysis has two parts. The first is a bound showing that force change per scan is bounded by the change in posterior mean and standard deviation, which in turn are bounded by kernel smoothness. The second is an empirical sweep of $m$ against latency and success, which gives the compute budget.

Baselines

The naive reference point is the standard ROS move_base stack with the Dynamic Window Approach (a local planner that samples velocity commands and scores them against a costmap), run with default inflation. The BARN 2022 report says all finalists "significantly outperformed" this baseline but gives no number, so no reported result was found on Lacuna for DWA on BARN.

The second baseline is a raw-point potential field in the style of Penders et al., with ranges mapped directly to repulsion and the same global planner and attractive term as ours.

The strongest simple competitor is the same raw-point field with median plus temporal filtering of ranges and a fixed inflation radius. Its filter window and radius get a tuning budget equal to ours on the clean and noisy training worlds. This baseline is the most likely to match us, and the main comparison is against it.

The fourth baseline is a per-beam uncertainty field. Each beam's temporal range variance inflates that beam's obstacle point, which gives uncertainty awareness without a GP.

The strongest existing full systems are the physical BARN winners. The 2022 winner, a classical hierarchical planner from UT Austin, completed 8/9 physical trials. The 2024 winner, LiCS-KI, a behavior-cloned transformer with a geometric safety layer, completed 6 of 9 physical trials with average traversal times as low as 27 seconds. These systems are not public in the file, so they serve as physical reference points rather than re-run baselines. In simulation we re-run a global-path-plus-MPC local planner (model predictive control, which optimizes a short-horizon trajectory at each step), matching the MLDA_EEE design (5 physical successes, 79 s average, BARN 2024).

Two ablations isolate the ingredients. The GP with $\kappa=0$ plus tuned fixed inflation isolates the uncertainty term. A GPnn-style per-query neighbor field isolates the fixed-per-scan conditioning.

All methods share the global planner, the robot model, velocity limits, and CPU hardware. Every method gets the same number of tuning trials on a held-out training split of worlds and noise settings. Beating the filtered-inflation baseline means a statistically significant success-rate gain on the heterogeneous-noise test worlds with no worse traversal time. On uniform-noise worlds we expect parity, and the claim predicts it.

Why not just…?

Why not just tune a larger fixed inflation radius? This is the real threat. With spatially uniform noise, $\kappa\sigma$ is roughly constant and fixed inflation is equivalent, so we expect a tie there and will report it. The argument for the GP rests on uneven corruption. A large radius deadlocks the robot in narrow gaps next to well-observed boxes. A small radius collides with dark boxes that return few points. The $\kappa=0$-plus-inflation ablation, evaluated on worlds mixing both, settles this. If it matches the full method, the uncertainty claim is false.

Why not just estimate per-beam variance and skip the GP? Per-beam statistics describe beams that return. A dropout on a dark surface produces no return, so there is no point to inflate and the region reads as free. Only a model with spatial correlation raises uncertainty in the gap between surrounding observations. The per-beam baseline, tested on dropout-patch worlds versus Gaussian-jitter worlds, shows whether this distinction matters in practice.

Why not just train an end-to-end policy with domain randomization? The file's evidence goes against it on this benchmark. DRL-VO ranked first in BARN 2022 simulation (score 0.2415) but completed 2/9 physical trials. The 2024 winner needed a hard-coded geometric safety layer. Our method can serve as that safety layer. We include a learned policy only as a secondary comparison, not as the main contest.

Why not just use EGO-Planner-style deterministic gradients? EGO-Planner reports about 0.81 ms planning time and about 89% success, but these are for quadrotors with crisp voxelized obstacles. Its anchors inherit every phantom and dropout. The raw-point field baseline plays this role on a ground robot.

Proposed Experiments and Decisive Checks

The main testbed is the 300 BARN Gazebo worlds (Jackal, 2D LiDAR), split into training and test sets. Each test world is run under four corruption regimes. The first is clean. The second adds uniform Gaussian range noise. The third assigns a subset of boxes a "dark" material that drops a set fraction of returns. The fourth injects isolated phantom returns at a set rate. The third and fourth regimes are also combined.

Metrics are success rate (goal reached without collision), collision rate, deadlock rate (no progress for a fixed interval), traversal time, command jitter (mean squared change in angular velocity between control steps), and per-cycle compute. No reported result was found on Lacuna for any method on noise-injected BARN, so all simulated numbers will be produced fresh under identical conditions.

Three checks decide the claim. The first is the $\kappa=0$ versus $\kappa>0$ gap: it should be near zero under uniform noise and clearly positive under dropout and phantom regimes. The second is the per-beam baseline against the GP on dropout worlds specifically. The third is jitter of the fixed-per-scan GP against the per-query-neighbor variant, where we predict the per-query variant shows force spikes and higher jitter.

A secondary check reuses the CARE setup. Monocular depth points replace LiDAR as GP input on a real TurtleBot4 corridor task. CARE reports up to 100% collision reduction and a 100% goal-reaching rate for ViNT+CARE on TurtleBot4, so the relevant question is whether our field cuts CARE's false avoidance maneuvers on glass and reflective obstacles without losing that rate. If a physical Jackal is available, a small cardboard-box maze with deliberately dark boxes gives a sim-to-real check against the physical reference numbers above.