NeSAM Boosts Off-Road Robot Motion Prediction Using Soil Physics
A neuro-symbolic model blends classical terramechanics with deep learning to improve trajectory tracking on deformable terrain.
Researchers have introduced NeSAM, a neuro-symbolic framework designed to improve how off-road vehicles predict their own motion over soft, deformable terrain such as sand or mud. As reported in a new paper posted to arXiv (cs.RO), the system combines physics-based soil modeling with machine learning to make long-horizon motion predictions more accurate and more interpretable than existing purely data-driven approaches.
Off-road mobility is notoriously difficult to model because soil conditions change constantly, altering sinkage, slip, and traction from one patch of ground to the next. According to the paper, existing learning-based kinodynamic models — which try to predict how a vehicle's motion and forces evolve over time — approximate these vehicle-terrain interactions directly from data but don't explicitly represent the underlying soil mechanics, limiting how interpretable and generalizable they are.
NeSAM tackles this by pairing a differentiable version of Bekker-Wong terramechanics theory, a classical framework for modeling vehicle-soil interaction forces, with learned terrain representations and a Transformer-based residual dynamics model. The terramechanics component estimates soil-dependent interaction forces, while the residual model corrects the gap between the analytical prediction and the vehicle's actual observed behavior. Together, the pipeline predicts six-degree-of-freedom kinodynamics, meaning it accounts for a vehicle's translational and rotational motion in three-dimensional space over long time horizons.
A distinguishing feature of the approach is that it estimates physically meaningful soil parameters directly from terrain observations, then continuously updates those estimates online using an extended Kalman filter as conditions shift during operation.
The team evaluated NeSAM in Verti-Bench, a simulator built on the Chrono multiphysics engine, and validated the results on a physical platform called Verti-4-Wheeler. Compared with the strongest baseline methods tested, NeSAM improved prediction accuracy by up to 30% in simulation and 29% on real-world data. When integrated with a closed-loop navigation controller, the system also improved traversal success rate through online soil adaptation, and reduced the Hausdorff distance to the reference trajectory — a measure of trajectory-tracking accuracy — by 69.4%.
The paper was submitted by Chenhui Pan on August 21, 2026, and is listed under the Robotics (cs.RO) category on arXiv.
talk to a robot
Wrong facts, cryptic artwork, missed news — pick the robot in charge and tell it directly. Litmus verifies the facts; what lands gets a thank-you engraved here.