From Simulation to Field: Learning Terrain Traversability for Real-World Deployment

TL;DR

This work tackles the problem of estimating traversability in unstructured outdoor settings with a direction-aware, continuous cost approach. It introduces TraverseNet, a LIDAR–IMU‑based neural network that yields robot‑direction sensitive traversability by fusing dense point clouds with inertial information, trained through an automated high‑fidelity simulation data pipeline and evaluated on both simulated and real platforms. Key contributions include automatic data generation for labeling via locomotion cues, dense robot-centric map construction using GPU-ICP, and demonstrations that the learned traversability maps improve path planning and autonomous exploration, even without real-world training data. The approach achieves state-of-the-art or superior results in MAE metrics on traversability costs, generalizes to real-world forest-like environments, and is released as open-source to accelerate practical deployment in field robotics.

Abstract

The challenge of traversability estimation is a crucial aspect of autonomous navigation in unstructured outdoor environments such as forests. It involves determining whether certain areas are passable or risky for robots, taking into account factors like terrain irregularities, slopes, and potential obstacles. The majority of current methods for traversability estimation operate on the assumption of an offline computation, overlooking the significant influence of the robot's heading direction on accurate traversability estimates. In this work, we introduce a deep neural network that uses detailed geometric environmental data together with the robot's recent movement characteristics. This fusion enables the generation of robot direction awareness and continuous traversability estimates, essential for enhancing robot autonomy in challenging terrains like dense forests. The efficacy and significance of our approach are underscored by experiments conducted on both simulated and real robotic platforms in various environments, yielding quantitatively superior performance results compared to existing methods. Moreover, we demonstrate that our method, trained exclusively in a high-fidelity simulated setting, can accurately predict traversability in real-world applications without any real data collection. Our experiments showcase the advantages of our method for optimizing path-planning and exploration tasks within difficult outdoor environments, underscoring its practicality for effective, real-world robotic navigation. In the spirit of collaborative advancement, we have made the code implementation available to the public.

Figures (18)

• Figure 1: Comprehensive Traversability Estimation Pipeline: Utilizing sensor inputs and high-fidelity simulations, data is processed via the EKF State Estimator and fed into TraverseNet. This deep neural network, trained on autonomously generated datasets, outputs precise traversability estimates vital for outdoor autonomous navigation components such as planning, control, and costmap.
• Figure 2: Fundamental TraverseNet Design: This illustration presents the core architecture, utilizing point coordinates data for feature extraction through convolutional and fully connected layers, subsequently integrated with a 13-dimensional IMU feature vector.
• Figure 3: Optimal TraverseNet Configuration: Detailed in Section \ref{['sec:Experiments']}, the architecture utilizing point coordinates (x, y, z) and point curvature inputs exhibits notable performance. Unlike design with direct IMU fusion depicted in \ref{['fig:base_traversenet']}, this approach processes IMU data through additional feature extraction before concatenating it with LIDAR features in the final layer.
• Figure 4: An illustration depicting the robot's computation of traversability labels using both nominal and actual travel distances. The blue represents the robot's actual path, the gold symbolizes the anticipated nominal path, and the green displays a time-series representation of the computed IMU acceleration covariance.
• Figure 5: A sequence of snapshots taken from a high-fidelity forest environment during a robot's data collection journey. Starting at the upper left, the robot progresses to its final position in the bottom right, where it encounters a non-traversable rock, marking the conclusion of a data collection "episode". The locomotion-based traversability, as discussed earlier, is implemented. Data samples are taken at intervals, ensuring the robot covers at least 1 meter between each data snapshot, aligning with the traversability label.