Paper Proposal: Terramechanics-Informed State Estimation for Tracked Autonomous Rovers

Proposed Title: Terramechanics-Informed State Estimation for Tracked Autonomous Rovers: Integrating Direct Visual-Inertial Odometry and 23-State Adaptive Fusion in Gazebo Target Venue: Journal of Field Robotics (JFR) or IEEE Robotics and Automation Letters (RA-L)

Authors & Affiliations

  • Khalid Bourr (Corresponding Author): Mechatronics & Robotics Lab, University of Trieste, Italy (unitsSpaceLab).

  • Manan Kharwar: Independent Researcher, Hamilton, ON, Canada.

  • Suyash: Core Open-Source Maintainer.

Abstract

Autonomous tracked rovers operating in agricultural and forestry environments suffer from severe localization drift due to track-soil slip. Standard wheel odometry fails in deformable terrain (mud, loose soil), and visual-inertial odometry (VIO) degrades under heavy canopy occlusion or low-angle glare. This paper presents a cascaded, terramechanics-aware sensor fusion framework deployed on a tracked agricultural robot (“Sowbot”). By integrating a modernized ROS 2 direct VIO pipeline with a 23-state Unscented Kalman Filter (UKF), the system dynamically down-weights slipping track encoders in favor of visual velocity during soil shearing events. To validate the architecture across the “reality gap,” we introduce a comprehensive simulation pipeline using Forest3D and gz-terramechanics, allowing for the formal verification of the fusion software against genuine, non-linear track-soil interaction models in Gazebo prior to field deployment.

1. Introduction and Motivation

Tracked vehicles in agriculture require centimeter-level precision for tasks like automated seeding, yet rely on tracks that inherently slip and shear against the topsoil to steer. Traditional navigation stacks treat the ground as a rigid 2D plane. When the Sowbot traverses wet clay, track encoders report forward motion while the vehicle is actually digging into the soil.

This paper solves this reality gap through a three-pillar architecture: a high-fidelity terramechanics simulation engine for validation, a local visual-inertial tracking engine, and a 23-state adaptive filter that cross-references physical slip against optical flow.

2. System Architecture

The proposed stack bridges high-rate VIO with adaptive state estimation, validated through procedural physics simulations.

Contributor

Subsystem

Function in the Stack

Khalid Bourr

Terramechanics Simulation

Leverages Forest3D to build complex DEM-based Gazebo worlds and gz-terramechanics to simulate physical sinkage, traction, and resistive torques. Generates the slip data that triggers the filter’s adaptive gating.

Suyash

Local VIO Engine

Maintains the modernized ROS 2 ROVIO implementation. Processes high-rate camera and IMU data to produce a clean velocity vector (/rovio/odometry) independent of ground friction.

Manan Kharwar

23-State UKF Fusion

Utilizes FusionCore to fuse VIO, IMU, and track data. Dynamically isolates the systematic yaw rate bias of the slipping tracks and coasts on VIO/IMU tracking when the physical tracks lose traction.

3. Experimental Validation in Gazebo

To prove the framework without risking physical hardware, the system is rigorously tested in a simulated high-slip environment:

  1. Environment Generation: High-resolution digital elevation models (DEM) of agricultural fields are processed using Forest3D. The resulting Gazebo world is populated with procedural vegetation to create realistic visual occlusions.

  2. Physics Injection: The gz-terramechanics plugin is applied to the soil. As the simulated Sowbot executes skid-steer maneuvers, the tracks calculate real-time sinkage and shear, causing the simulated wheel encoders to spike out of sync with actual displacement.

  3. Filter Intercept: FusionCore monitors the divergence between the VIO velocity and the track encoders. As the slip crosses the Mahalanobis chi-squared threshold, the filter drops the encoder weights, relying entirely on the visual map and IMU dead-reckoning to maintain an accurate trajectory.

4. Bibliography and Prior Work

The foundational frameworks enabling this research are rooted in the authors’ prior active developments in state estimation and simulation:

FusionCore & State Estimation (Manan Kharwar)

  • Kharwar, M. (2026). FusionCore: A 23-State Unscented Kalman Filter for IMU, Wheel Encoder, GPS, and Visual SLAM Fusion in ROS 2. arXiv preprint arXiv:2605.25239. [Details the core UKF math, online bias estimation, and adaptive noise covariance used to intercept the VIO data].

Terramechanics & Environment Generation (Khalid Bourr)

  • Scalera, L., Maset, E., Fasiolo, D. T., Bourr, K., Cottiga, S., De Lorenzo, A., Carabin, G., Alberti, G., Gasparetto, A., Mazzetto, F., & Seriani, S. (2026). Forest Surveying with Robotics and AI: SLAM-Based Mapping, Terrain-Aware Navigation, and Tree Parameter Estimation. Machines.

  • Bourr, K. (2026). Forest3D - Terrain and Forest Generation for Gazebo. unitsSpaceLab, GitHub. [Automated generation of Gazebo environments from DEM files and Blender assets].

  • Bourr, K. (2026). gz-terramechanics. unitsSpaceLab, GitHub. [Gazebo Sim physics plugin computing real-time wheel/track sinkage, traction forces, and driving torques].

Visual-Inertial Odometry (Suyash & ETH Zürich)

  • Bloesch, M., Omari, S., Hutter, M., & Siegwart, R. (2015). Robust visual inertial odometry using a direct EKF-based approach. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). [The foundational math modernized into ROS 2 by Suyash].