A Self-Supervised Terrain Roughness Estimator for Off-Road Autonomous Driving (1206.6872v1)

Abstract: We present a machine learning approach for estimating the second derivative of a drivable surface, its roughness. Robot perception generally focuses on the first derivative, obstacle detection. However, the second derivative is also important due to its direct relation (with speed) to the shock the vehicle experiences. Knowing the second derivative allows a vehicle to slow down in advance of rough terrain. Estimating the second derivative is challenging due to uncertainty. For example, at range, laser readings may be so sparse that significant information about the surface is missing. Also, a high degree of precision is required in projecting laser readings. This precision may be unavailable due to latency or error in the pose estimation. We model these sources of error as a multivariate polynomial. Its coefficients are learned using the shock data as ground truth -- the accelerometers are used to train the lasers. The resulting classifier operates on individual laser readings from a road surface described by a 3D point cloud. The classifier identifies sections of road where the second derivative is likely to be large. Thus, the vehicle can slow down in advance, reducing the shock it experiences. The algorithm is an evolution of one we used in the 2005 DARPA Grand Challenge. We analyze it using data from that route.

An Analysis of Self-Supervised Terrain Roughness Estimation for Autonomous Off-Road Driving

The paper "A Self-Supervised Terrain Roughness Estimator for Off-Road Autonomous Driving" by David Stavens and Sebastian Thrun presents a novel machine learning approach suitable for off-road autonomous vehicle perception, focused on real-time terrain roughness estimation using laser range data. This self-supervised method is specifically designed to enhance the vehicle’s capability to navigate rough terrains and reduce shock impacts that occur at high speeds. The research aims for greater precision than traditional obstacle detection by leveraging the data from the 2005 DARPA Grand Challenge.

Methodology

The authors propose a self-supervised learning method that generates training labels autonomously from the vehicle's own inertial measurements. The ruggedness of the terrain is derived by comparing sets of nearby laser-measured surface points. This task is complicated by uncertainties such as sparse data collection at distance and inaccuracies in laser point projection caused by latency and pose estimation errors. To address these challenges, the research models these errors as multivariate polynomials whose coefficients are optimized through self-supervised learning.

The terrain roughness "labels" are automatically created from the shock data measured by the vehicle's inertial sensors as it traverses the terrain. This transference of roughness detection capability from inertial sensors to laser sensors allows preemptive speed control based on terrain roughness predictions rather than reactive speed adjustments post-shock detection.

Experimental Results

This method was evaluated using data from the DARPA Grand Challenge, specifically a segment between miles 60-70 for training and miles 70-80 for testing. It was found that this approach more effectively identifies rough surfaces and ensures vehicle safety, achieving up to a 50% reduction in vehicle shock experiences compared to prior methods.

The Receiver Operating Characteristic (ROC) curve analysis reveals that this method achieves a higher true-positive rate across most false-positive thresholds compared to previous obstacle detection algorithms. Consequently, vehicles can anticipate rough terrain and adjust speed proactively, a significant improvement over the reactive methods used during the actual DARPA Grand Challenge event. The self-supervised method offers substantial improvements in reducing vehicle shock without considerably extending travel time.

Implications and Future Work

The implications of this research extend to both theoretical and practical domains in the field of autonomous vehicle navigation. Theoretically, it introduces an efficiently self-supervised approach that bypasses the need for manually labeled data, providing a framework that can potentially be applied in various domains requiring real-time environmental assessment. Practically, the reduction in vehicle shock directly translates into improved vehicular longevity and passenger comfort in autonomous vehicles.

Future developments may focus on integrating vision systems to extend rough terrain detection ranges beyond those feasible with only lidar data. Refinement of the mathematical model used for surface analysis could enhance performance accuracy. Lastly, transitioning the approach from an offline to an online process could improve adaptability and robustness during real-time navigation.

By investigating these avenues, the research could contribute significantly to the development of fully autonomous vehicles capable of safely navigating diverse and unpredictable off-road environments.