How Does It Feel? Self-Supervised Costmap Learning for Off-Road Vehicle Traversability (2209.10788v3)

Abstract: Estimating terrain traversability in off-road environments requires reasoning about complex interaction dynamics between the robot and these terrains. However, it is challenging to create informative labels to learn a model in a supervised manner for these interactions. We propose a method that learns to predict traversability costmaps by combining exteroceptive environmental information with proprioceptive terrain interaction feedback in a self-supervised manner. Additionally, we propose a novel way of incorporating robot velocity in the costmap prediction pipeline. We validate our method in multiple short and large-scale navigation tasks on challenging off-road terrains using two different large, all-terrain robots. Our short-scale navigation results show that using our learned costmaps leads to overall smoother navigation, and provides the robot with a more fine-grained understanding of the robot-terrain interactions. Our large-scale navigation trials show that we can reduce the number of interventions by up to 57% compared to an occupancy-based navigation baseline in challenging off-road courses ranging from 400 m to 3150 m. Appendix and full experiment videos can be found in our website: https://mateoguaman.github.io/hdif.

• The paper presents a self-supervised technique that fuses high-dimensional visual and geometric data with IMU feedback to generate continuous traversability costmaps without human labels.
• It incorporates robot velocity via Fourier feature mapping, enabling dynamic cost predictions that account for terrain and speed-dependent forces.
• The method integrates learned costmaps with traditional path planning, demonstrating smoother navigation and a 57% reduction in intervention events on diverse off-road platforms.

Self-Supervised Costmap Learning for Off-Road Vehicle Traversability

The paper "How Does It Feel? Self-Supervised Costmap Learning for Off-Road Vehicle Traversability" presents a novel approach that combines exteroceptive environmental data and proprioceptive feedback to predict traversability costmaps for off-road robots in a self-supervised manner. Off-road environments pose significant challenges due to complex interactions between the robot and terrain, which traditional occupancy-based navigation methodologies fail to address adequately. The authors propose a method that not only captures these interactions but also introduces robot velocity into the prediction pipeline, thus allowing the robot to reason about the traversability at different speeds.

Key Contributions

1. Self-supervised Learning for Costmap Prediction: The paper introduces a method to predict continuous traversability costmaps using a combination of high-dimensional visual and geometric data with proprioceptive feedback. The costmaps are learned without requiring human-annotated labels, using a pseudo-ground truth derived from inertial measurement unit (IMU) data.
2. Velocity Incorporation: The inclusion of robot velocity in the prediction model through Fourier feature mapping distinguishes this approach. This consideration is crucial as dynamic forces in robot-terrain interactions are velocity-dependent.
3. Effective Integration with Navigation Systems: The proposed method demonstrates improved navigation performance by integrating learned costmaps with traditional path planning and control elements, demonstrating reduced interventions and smoother navigation on varied terrains.
4. Generalization Across Robots: The system is demonstrated on two separate large ground robots—an all-terrain vehicle (ATV) and a Clearpath Warthog unmanned ground vehicle (UGV)—showcasing its adaptability and transferability across different robotic platforms.

Methodology

• Traversability Cost Derivation: Utilizing the power spectral density of vertical linear acceleration from IMU data, the authors derive a continuous traversability cost function. This pseudo-ground truth captures terrain interaction dynamics and is computed in the 1-30 Hz frequency range, selected based on empirical correlations with human-labeled roughness scores.
• Visuospatial and Geometric Mapping: The use of bird's-eye-view maps aggregates visual RGB and geometric information from stereo camera disparity images, facilitating the representation of environmental features in a format conducive to learning.
• Neural Network Architecture: The architecture integrates a convolutional neural network (CNN) backbone for processing visual inputs with a multi-layer perceptron (MLP) that handles the Fourier parameterized velocity. This combination allows seamless fusion of high and low-dimensional data inputs to generate the final cost predictions.

Evaluation and Results

The paper provides an empirical assessment of the proposed system across various navigation tasks, validating the effectiveness of the predicted costmaps through both short-scale and large-scale navigation experiments. Notably, in large-scale navigation courses, the proposed approach resulted in a 57% reduction in intervention events compared to occupancy-based methods. By effectively capturing terrain nuances like smoothness and roughness, the learned costmaps guide these robots to choose smoother navigational paths, thereby enhancing autonomy in challenging off-road environments.

Implications and Future Directions

The concept of self-supervised learning for traversability prediction opens new avenues for enhancing the autonomous capabilities of off-road robots. By enabling these systems to perceive and adapt to complex terrains dynamically, this methodology lays the groundwork for deeper integration of machine learning with robot navigation stacks. Potential future developments could include improving real-time processing capabilities, exploring online adaptation mechanisms for different environments, and extending this approach to other robotic navigation challenges. Additionally, further refinement of feature representation methods might address current limitations in perception accuracy and system latency. Overall, this work significantly contributes to the practical advancement of autonomous robotics in unstructured environments.