Vehicle-Agnostic Terrain Posteriors from Simulation-Based Inference

Most off-road robots still decide where to drive with a map that marks each patch of ground as either passable or blocked. That picture fails in two directions. Mud hidden under grass gets marked passable and the vehicle gets stuck. Soft brush gets marked as a wall and the vehicle takes a long detour. Learned costmaps have replaced the binary map with a continuous score. But that score is learned from the logs of one vehicle, so it describes how that particular vehicle felt the ground, not what the ground is. A score learned on a heavy all-terrain vehicle says little about a light rover with narrow wheels.

We propose to estimate the physical properties of the ground itself, with error bars, and then compute each vehicle's risk from those properties. The ground properties we mean are soil stiffness, how quickly soil gives way under shear, and how deep wheels sink. The tool is simulation-based inference: a family of methods that fit the parameters of a stochastic simulator to real observations even when the probability of an observation cannot be written down. Here it is used in two steps. First, a network trained only on simulated drives reads a few seconds of the vehicle's motion sensors and returns a probability distribution over the soil parameters of the patch just driven over. Second, those distributions become training targets for a camera-and-LiDAR network that predicts a soil-parameter distribution for ground the vehicle has not yet reached. A planner then pushes samples from that distribution through any vehicle's fast wheel-soil model to estimate that vehicle's tail risk.

Core Claim

The claim is that physically grounded, uncertain soil parameters transfer across vehicles, and capture rare dangerous terrain, better than costs learned directly from proprioception. Proprioception here means the vehicle's internal sensing: accelerometers, wheel speed and suspension travel. On the vehicle that collected the data and on terrain it drove often, we expect no advantage over direct cost learning, and we do not claim one. Where we expect the method to win is in predicting outcomes for a second vehicle with different wheels or mass. We also expect it to win on the worst few percent of events, such as deep sinkage, high slip and near-entrenchment. Driving logs rarely contain those events because human drivers avoid them, while a simulator can generate them on demand. If neither advantage appears, the simulator is not earning its cost, and the proposal should be judged a failure.

Relation to Prior Work

The closest work applies likelihood-free inference to robot physics. BayesSim learns a mixture-of-Gaussians posterior over global simulator constants, such as friction and mass, from hand-designed trajectory summaries. It uses those posteriors to set the parameter ranges a policy is trained under, which is called domain randomization. Inferring the Material Properties of Granular Media wraps a GPU particle simulator in a BayesSim-style loop and recovers grain friction from 16 hand-built depth statistics of poured piles, with a sim-to-sim sliding-friction error of 0.0062. Both papers supply our method and show that the inference step works for contact physics. Both infer one global constant for a static lab scene. We infer a parameter distribution for every map cell, from moving-vehicle signals, and then pass it ahead of the vehicle through perception.

The articulated wheeled bevameter shapes the design in a different way. It argues that vision alone cannot reveal soil firmness, and it measures pressure-sinkage and shear with a dedicated probing wheel on a robot arm. We accept the argument and replace the arm with the drive wheels: every traversed patch becomes a probe. We also accept that a vision-only prediction should be a wide prior, not a confident estimate.

Three lines of work learn continuous traversability from real data and are our main competitors. How Does It Feel? uses vertical-acceleration bandpower as a self-supervised cost. Learning Risk-aware Costmaps predicts tail quantiles of cost, specifically Conditional Value-at-Risk (CVaR): the average cost of the worst fraction of outcomes. EVORA predicts a Dirichlet distribution over traction and flags out-of-distribution inputs with a normalizing flow. All three model the risk of one platform. None of them ties uncertainty to physical degrees of freedom that could be re-used for another vehicle.

Two further papers matter for specific steps. Hybrid dynamics models, such as the DINOv2-plus-bicycle-model network of Gibson et al., learn force corrections that are tied to one vehicle. The concern that standard simulation-based posteriors become overconfident when the simulator is wrong comes from Misspecification-robust Sequential Neural Likelihood. The tempered fix we adopt comes from Amortized SBI in Generalized Bayes via NPE, which so far has only been tested on synthetic tasks.

Research Program

Simulator. We build a stochastic wheel-soil simulator: a multibody vehicle model whose wheels interact with the ground through Bekker-Wong pressure-sinkage and Janosi-Hanamoto shear laws. Per-patch soil parameters are drawn from a prior. To make the likelihood intractable in the way real ground is, the simulator also randomizes hidden micro-terrain: buried rocks, moisture-dependent changes in cohesion, and vegetation drag. Vehicle geometry is an explicit input, so one simulator covers the TartanDrive Yamaha Viking, a smaller skid-steer robot, and others.

Step 1: a proprioceptive posterior estimator. We train a neural posterior estimator, a conditional normalizing flow $q_\phi(\theta \mid w)$. Here $\theta$ is the soil-parameter vector of a patch, and $w$ is a short window of accelerometer readings, wheel speed, shock travel, and commanded versus achieved speed recorded while crossing that patch. Following the generalized-Bayes approach, the posterior is tempered by a factor $\beta$. We choose $\beta$ so that coverage is correct on held-out simulations from a deliberately mismatched simulator variant, for example one with fibrous vegetation added. Because the inputs are a few low-dimensional time series rather than images, the main dimensionality problem of simulation-based inference does not arise in this step. This is a deliberate design choice.

Step 2: a forward-looking prior. We run $q_\phi$ over real TartanDrive logs. Every traversed cell receives a posterior distribution over soil parameters. A bird's-eye-view perception network is then trained to output a distribution $r_\psi(\theta \mid v)$, where $v$ is the camera and LiDAR view of a cell before the vehicle reaches it. Training minimizes $\mathbb{E}_{\theta \sim q_\phi(\cdot \mid w)}[-\log r_\psi(\theta \mid v)]$. In words, the visual network learns to cover the whole spread of soil parameters that proprioception allows, not a single point value.

Step 3: planning. For any vehicle, cell cost is the CVaR of sinkage, slip, or energy. It is computed by drawing samples of $\theta$ from $r_\psi$ and running that vehicle's fast single-wheel model on each sample. When the vehicle drives over a cell, $q_\phi$ replaces the prior for that cell.

Baselines

The naive reference is a geometric costmap built from slope and roughness with Bayesian Generalized Kernel inference. It fills in sparse LiDAR to produce height maps and traversability probabilities at 10 Hz. It shows what can be done without any physical information about the soil.

The strongest competitor is a direct real-data risk model in the style of EVORA and Fan et al.. It is trained on the same bird's-eye-view inputs and the same TartanDrive logs to predict quantiles of slip and vertical-acceleration bandpower directly, with no simulator. It is the strongest competitor because it is exactly what a practitioner would build. On the platform that collected the data it is plausibly as good as anything else.

For the published reference point, TartanDrive reports a state-prediction RMSE of 0.1674 for its full multimodal model, against 1.1638 for a kinematic baseline. We do not have the competitor's numbers for tail calibration or cross-platform prediction: no reported result was found on Lacuna for EVORA or the risk-aware costmap on TartanDrive.

Fairness rules apply to every comparison. All methods use the same perception backbone, the same training split, and the same hyperparameter search budget. The simulator compute we spend is reported separately, so readers can weigh it.

Beating the competitor means one of two things on a second vehicle, with the comparison held fixed. Either our method has lower negative log-likelihood and better 90%-interval coverage on held-out slip and sinkage, or it has a smaller error on the worst-5% events. Same-platform parity is the expected result and does not count as a win.

Three ablations isolate the key ingredients. Removing the posterior replaces $q_\phi$ with a least-squares point fit of Bekker-Wong parameters per patch. Removing the physics replaces $\theta$ with an unconstrained learned latent of the same size. Removing the inference trains the planner on uniformly randomized parameters, which is the comparison BayesSim used.

Why not just…?

Why not just learn risk directly from real proprioception, as How Does It Feel? and EVORA do? This is the real threat. For the logging vehicle it may well match us; How Does It Feel? already reports up to 57% fewer interventions than an occupancy baseline over trials of up to 3150 m. The cross-vehicle and tail experiments below exist to settle the question. A learned cost encodes one vehicle's response. To use it for a second vehicle you need that vehicle's data, and rare failures are thin in any log. If the direct model transfers just as well, we will report that the simulator adds nothing.

Why not just fit Bekker-Wong equations by regression? The bevameter study shows that these equations are highly sensitive to parameter error. A point fit also hides the fact that many different soils produce the same wheel signal. The point-fit ablation tests this directly. If it matches the posterior on tail-risk error, the posterior machinery is unnecessary.

Why not just randomize simulator parameters widely and skip inference? BayesSim found that uniform randomization produced higher-variance, lower-reward policies than inferred posteriors on Fetch Slide. We repeat that comparison for planning on deformable ground through the uniform-randomization ablation.

Proposed Experiments

Simulation tests. In simulation we check posterior calibration using simulation-based calibration rank statistics. We do this on the matched simulator and on mismatched variants, with and without tempering. The goal is to confirm that tempering restores coverage without making posteriors uselessly wide.

Real-data prediction. On TartanDrive and TartanDrive 2.0, we predict proprioceptive outcomes 5 to 30 m ahead of the vehicle on held-out trajectories. We report negative log-likelihood, interval coverage, and error on the worst 5% of events. One known risk: heavy slip corrupts odometry, which TerrainNet notes degrades mapping. We therefore compute achieved speed from multiple fused LiDARs, not from wheel odometry.

Cross-vehicle test. This is the decisive experiment. We train every method on Yamaha Viking data only, then predict slip and sinkage for a second, lighter platform. The platform is a Clearpath Warthog, as used by Triest et al., driven over the same instrumented sites in a small field collection. Our method changes only the vehicle model inside the simulator. The direct-learning baseline must either be applied unchanged or be fine-tuned on a matched, small amount of Warthog data.

Closed-loop planning. In closed-loop simulation with procedurally hidden soft patches, we compare entrapment rate and path length against the geometric and direct-risk planners.

What counts as success. The proposal succeeds if, on the second vehicle, our method gives better-calibrated tail predictions than the direct-risk model, and the physics-free latent ablation loses that advantage. Same-platform parity combined with no cross-vehicle gain would falsify the core claim.