DRIVE Through the Unpredictability:From a Protocol Investigating Slip to a Metric Estimating Command Uncertainty

Abstract: Off-road autonomous navigation is a challenging task as it is mainly dependent on the accuracy of the motion model. Motion model performances are limited by their ability to predict the interaction between the terrain and the UGV, which an onboard sensor can not directly measure. In this work, we propose using the DRIVE protocol to standardize the collection of data for system identification and characterization of the slip state space. We validated this protocol by acquiring a dataset with two platforms (from 75 kg to 470 kg) on six terrains (i.e., asphalt, grass, gravel, ice, mud, sand) for a total of 4.9 hours and 14.7 km. Using this data, we evaluate the DRIVE protocol's ability to explore the velocity command space and identify the reachable velocities for terrain-robot interactions. We investigated the transfer function between the command velocity space and the resulting steady-state slip for an SSMR. An unpredictability metric is proposed to estimate command uncertainty and help assess risk likelihood and severity in deployment. Finally, we share our lessons learned on running system identification on large UGV to help the community.

• The paper introduces the DRIVE protocol, a standardized method for systematic UGV system identification and slip characterization.
• It proposes an unpredictability metric that quantifies command uncertainty by comparing commanded and measured kinetic energy.
• Empirical results over 4.9 hours and 14.7 km validate the protocol’s effectiveness in capturing slip behaviors and risk trends across varied terrains.

DRIVE: Protocolizing Slip Characterization and Command Uncertainty Estimation in Off-Road UGVs

The paper presents a systematic approach to data-driven system identification for Uncrewed Ground Vehicles (UGVs) navigating off-road, with focus on characterizing vehicle-terrain interactions, quantifying slippage, and assessing command uncertainty through a combination of experimental protocols and novel metrics. The authors propose and validate the DRIVE (Data-driven Robot Input Vector Exploration) protocol as a repeatable methodology for generating comprehensive datasets essential for understanding and mitigating slip-induced unpredictability in mobile robot control.

Motivation and Context

Accurate motion models underpin robust localization, planning, and control for autonomous navigation. Off-road conditions pose significant challenges: slippage arises from complex vehicle–terrain interactions poorly captured by standard exteroceptive or proprioceptive sensors. While system identification through empirical data collection is customary, lack of standardized, reproducible protocols leads to incomplete coverage of the command space, inconsistent datasets, and unreliable comparison between methods or platforms.

Existing data-gathering strategies cover only a subset of the possible vehicle state and input space, often overrepresenting benign driving behaviors or omitting aggressive maneuvers, terrain-induced disturbances, or states far from nominal conditions. This hampers the development and benchmarking of motion models, as well as objective assessment of risk and prediction uncertainty.

Protocol Definition and Implementation

The DRIVE protocol is introduced to systematize the exploration of a UGV's command space—in both wheel and body frames—to ensure broad coverage of possible input-output conditions over varying terrain types. Its implementation involves:

• Command Space Identification: Determining true vehicle actuation limits (e.g., velocity and acceleration bounds) through both specifications and empirical calibration, accounting for all system layers (high-level planners, low-level controllers, motor dynamics).
• Terrain Selection and Constraints: Choosing sufficiently flat, homogeneous terrain areas to isolate terrain–vehicle interaction effects, while recognizing residual variability in real-world field deployments.
• Randomized Uniform Command Sampling: Generating and executing random velocity (and optional acceleration) commands within bounds to sufficiently excite the system, for both transient and steady-state responses.
• Sensor Instrumentation and Estimation: Logging wheel encoder feedback, body velocities (typically via lidar/SLAM for linear motion, IMU for rotation), and all commanded inputs, synchronized at sufficient rates.
• Safety Monitoring: Utilizing remote operator intervention to handle out-of-bounds conditions or safety-critical states, ensuring protocol continuation and dataset consistency.

A detailed algorithmic procedure (see Algorithm 1 in the paper) and open-source implementation are provided to facilitate replication and extension for different UGV platforms.

Unpredictability Metric

A core contribution is the definition and empirical evaluation of an unpredictability metric $\rho$ to quantify command uncertainty: the divergence between commanded and realized kinetic energy (combining translational and rotational components via the inertia matrix). The ratio is penalized for vector misalignment, yielding a continuous, bounded measure ($[0,1]$) indicative of actuation uncertainty, energy loss, or slip severity.

The key attributes of this metric include:

• Consistency across different platforms and terrains.
• Sensitivity to both internal (vehicle, actuation limits) and external (terrain, environment) slip factors.
• Scalability and low implementation overhead (requiring only velocity measurements, robot geometry, and commanded inputs).
• Utility as both a deployment risk indicator and a comparative tool for empirical evaluation across studies.

Empirical Results

Experiments span over 4.9 hours and 14.7 km of driving with two SSMR platforms (Clearpath Husky and Warthog) across diverse terrains: asphalt, grass, gravel, ice, mud, and sand. Key findings include:

• Coverage: The DRIVE protocol effectively excites both the internal state-space (command, slip, and body velocities) and reveals terrain-imposed boundaries on reachable commands and velocities.
• Slip Characterization: Map-like transfer functions between commanded and measured velocities are estimated via Gaussian kernel smoothing, enabling explicit analysis of how different terrains modulate longitudinal, lateral, and rotational slip.
• Unpredictability Trends: Highly slippable terrains (ice, wet mud, sand) yield elevated unpredictability scores, with strong correlations between areas of command space and slip-prone behaviors (e.g., high angular and longitudinal speeds resulting in drift).
• Risk Assessment: The unpredictability metric, combined with vehicle kinetic energy, enables continuum-based risk assessment, moving beyond conventional categorical risk matrices. This approach more accurately reflects scenario-dependent severity and likelihood of navigation failures or hazardous events.

Implementation Considerations

• Computational Load: The method is lightweight in terms of online computation, as the main burden is dataset collection and offline analysis. Real-time calculation of the unpredictability metric is feasible for online risk assessment.
• Practical Limitations:
  • Mechanical and tire wear: Wide exploration of high-torque commands can accelerate hardware degradation, necessitating careful limits and environment selection.
  • Energy consumption: Aggressive maneuvers on deformable terrain increase battery usage; opting for uniform but not excessive sampling is advised.
  • Environmental impact: Repetitive aggressive driving can alter terrain properties, potentially invalidating the protocol's homogeneity assumption.
  • Localization: Accurate linear velocity (especially non-holonomic components) is more challenging to estimate, potentially affecting metric reliability.

Theoretical and Practical Implications

The standardization of data-gathering protocols, as advanced by DRIVE, enhances reproducibility, comparability, and benchmarking of motion models for robotics researchers. The unpredictability metric provides an objective means to characterize deployment environments and quantify risk, reducing reliance on subjective or human-curated terrain labels. These tools enable:

• Better transferability of empirical models across platforms and environments.
• More reliable field deployment risk analyses, both for research and industrial deployment (e.g., autonomous fleets, safety-critical systems).
• Informed benchmarking of SLAM/localization and control algorithms under challenging driving conditions, supporting both development and regulatory compliance.

Future Research Directions

• Protocol Generalization: Extending DRIVE to holonomic and articulated platforms, or to more complex multi-body UGVs.
• Active Sampling: Incorporating adaptive or GP-based exploration strategies to further increase efficiency and identification accuracy.
• Dynamic Terrain Modeling: Addressing terrain deformation effects and history-dependent properties in real-time slip identification and prediction.
• Real-time Control: Direct integration of the unpredictability metric into motion planners, enabling dynamic risk-aware navigation.

Strong Claims and Numerical Results

• The protocol enables exploration of the velocity state-space on multiple terrains, generating data that leads to slip model estimation across all significant motion dimensions.
• Certain terrains (notably ice) induce command unpredictability values over 1.6 times higher than combined values from hard grounds (grass, asphalt, gravel).
• Aggressive driving configurations (high angular and linear speeds together) achieve up to 50% of maximum possible lateral slip on ice, while being significantly suppressed on high-friction surfaces.

Conclusion

This work delivers a robust foundation for empirical system identification in off-road robotics, combining a replicable data collection methodology (DRIVE) with a continuous, physically-grounded unpredictability metric. These advances collectively address core challenges in risk quantification, reproducibility, and assessment of command uncertainty, with clear paths for practical uptake and ongoing methodological improvement in autonomous robotics research and application.