TRIP: Terrain Traversability Mapping With Risk-Aware Prediction for Enhanced Online Quadrupedal Robot Navigation

TL;DR

TRIP reconstructs the terrain maps while predicting multi-modal traversability risks, enhancing online autonomous navigation with the following contributions: estimating steppability in a spherical projection space allows for addressing data sparsity while accomodating scalable terrain properties.

Abstract

Accurate traversability estimation using an online dense terrain map is crucial for safe navigation in challenging environments like construction and disaster areas. However, traversability estimation for legged robots on rough terrains faces substantial challenges owing to limited terrain information caused by restricted field-of-view, and data occlusion and sparsity. To robustly map traversable regions, we introduce terrain traversability mapping with risk-aware prediction (TRIP). TRIP reconstructs the terrain maps while predicting multi-modal traversability risks, enhancing online autonomous navigation with the following contributions. Firstly, estimating steppability in a spherical projection space allows for addressing data sparsity while accomodating scalable terrain properties. Moreover, the proposed traversability-aware Bayesian generalized kernel (T-BGK)-based inference method enhances terrain completion accuracy and efficiency. Lastly, leveraging the steppability-based Mahalanobis distance contributes to robustness against outliers and dynamic elements, ultimately yielding a static terrain traversability map. As verified in both public and our in-house datasets, our TRIP shows significant performance increases in terms of terrain reconstruction and navigation map. A demo video that demonstrates its feasibility as an integral component within an onboard online autonomous navigation system for quadruped robots is available at https://youtu.be/d7HlqAP4l0c.

Figures (8)

• Figure 1: Overview of terrain traversability mapping with risk-aware prediction (TRIP). TRIP generates a local and global terrain map with multi-modal traversability risk prediction, enhancing online quadruped robot navigation. All figures in this paper are best viewed in color.
• Figure 2: Overview of the proposed TRIP framework. (a) Using 3D LiDAR or depth camera, a surfel map $\mathcal{S}$ is generated from a range map. And the surfel map is used to build a steppability risk maps $\tilde{\mathcal{I}}^{r}$ and $\mathcal{I}^{r}$. (b) After projecting the surfel and the steppability risk onto the elevation map $\mathcal{E}$, the local terrain map $\bar{\mathcal{E}}$ is completed based on traversability-aware Bayesian generalized kernel (T-BGK) $k^{\mathcal{T}}$ and inference function $\mathcal{L}^{\mathcal{T}}$ while embedding traversability risks. (c) Outliers are rejected using steppability-based Mahalanobis distance, and a static terrain map $\hat{\mathcal{E}}$ is updated with our proposed bias models. Each risk layer in the terrain map is color-coded to represent varying risk levels. Yellow, orange, and green shades indicate low risk levels, while purple, blue, and black denote high-risk areas. This color scheme is consistently employed for terrain map figures throughout this paper.
• Figure 3: Steppability map results from (a) surfel map $\mathcal{S}$ from the range sensor with $1\text{~pixel}=1^\circ \times 1^\circ$ resolution: (b) Raw risk map $\tilde{\mathcal{I}}^{r}$, refined risk maps with (c) our conditional pooling $\mathcal{I}^{r}$. Our conditional pooling successfully reduces noise and enhances the risk discernment.
• Figure 4: Local inference results, $\bar{\mathcal{E}}$, on the enclosed flat and bumpy terrains: (a) Vanila BGK-based inference $\mathcal{L}(\cdot)$ results. (b) The proposed T-BGK-based inference $\mathcal{L}^{\mathcal{T}}(\cdot)$ results. The grey points represent the predicted results of each inference function. Vanila BGK infers unobservable regions beyond walls (red areas) and uses every neighbors without any consideration of steppability when predicting the empty cells (cyan areas). In contrast, the proposed T-BGK distinguishes observable regions and is based on $r^{\mathrm{step}}$, allowing for traversability-aware predictions.
• Figure 5: Example of outliers and sequences of our steppability-based Mahalanobis distance filtering in a real-world environment. (a) The surfel map and the steppability risk map. Among (b) the local terrain maps which are completed for each time, (c) the outliers are filtered based on the updated terrain map at the previous time. (d) After the rejection, the terrain map is updated based on Kalman filter. The red points represent the traces of rejected outliers. Our proposed rejection-based map update module enhances robustness of our terrain map against dynamic elements and sensor noises.