Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

UNRealNet: Learning Uncertainty-Aware Navigation Features from High-Fidelity Scans of Real Environments

Samuel Triest, David D. Fan, Sebastian Scherer, Ali-Akbar Agha-Mohammadi

TL;DR

UNRealNet tackles uncertainty-aware traversability in rugged, unstructured environments by learning dense navigation features from high-fidelity real-world scans and predicting them from sparse lidar inputs on robots. The approach combines a high-quality data pipeline with a PointPillars-based network that outputs per-cell Gaussian feature distributions and a probabilistic traversability function that leverages uncertainty for robot-specific planning. Empirical results show improvements over traditional baselines in both synthetic and real hardware evaluations, including better handling of occlusions and availability of robot-agnostic predictions. The method demonstrates practical viability on legged platforms, with a Kalman-filter fusion to further enhance feature estimates and feasible runtime for on-robot deployment.

Abstract

Traversability estimation in rugged, unstructured environments remains a challenging problem in field robotics. Often, the need for precise, accurate traversability estimation is in direct opposition to the limited sensing and compute capability present on affordable, small-scale mobile robots. To address this issue, we present a novel method to learn [u]ncertainty-aware [n]avigation features from high-fidelity scans of [real]-world environments (UNRealNet). This network can be deployed on-robot to predict these high-fidelity features using input from lower-quality sensors. UNRealNet predicts dense, metric-space features directly from single-frame lidar scans, thus reducing the effects of occlusion and odometry error. Our approach is label-free, and is able to produce traversability estimates that are robot-agnostic. Additionally, we can leverage UNRealNet's predictive uncertainty to both produce risk-aware traversability estimates, and refine our feature predictions over time. We find that our method outperforms traditional local mapping and inpainting baselines by up to 40%, and demonstrate its efficacy on multiple legged platforms.

UNRealNet: Learning Uncertainty-Aware Navigation Features from High-Fidelity Scans of Real Environments

TL;DR

UNRealNet tackles uncertainty-aware traversability in rugged, unstructured environments by learning dense navigation features from high-fidelity real-world scans and predicting them from sparse lidar inputs on robots. The approach combines a high-quality data pipeline with a PointPillars-based network that outputs per-cell Gaussian feature distributions and a probabilistic traversability function that leverages uncertainty for robot-specific planning. Empirical results show improvements over traditional baselines in both synthetic and real hardware evaluations, including better handling of occlusions and availability of robot-agnostic predictions. The method demonstrates practical viability on legged platforms, with a Kalman-filter fusion to further enhance feature estimates and feasible runtime for on-robot deployment.

Abstract

Traversability estimation in rugged, unstructured environments remains a challenging problem in field robotics. Often, the need for precise, accurate traversability estimation is in direct opposition to the limited sensing and compute capability present on affordable, small-scale mobile robots. To address this issue, we present a novel method to learn [u]ncertainty-aware [n]avigation features from high-fidelity scans of [real]-world environments (UNRealNet). This network can be deployed on-robot to predict these high-fidelity features using input from lower-quality sensors. UNRealNet predicts dense, metric-space features directly from single-frame lidar scans, thus reducing the effects of occlusion and odometry error. Our approach is label-free, and is able to produce traversability estimates that are robot-agnostic. Additionally, we can leverage UNRealNet's predictive uncertainty to both produce risk-aware traversability estimates, and refine our feature predictions over time. We find that our method outperforms traditional local mapping and inpainting baselines by up to 40%, and demonstrate its efficacy on multiple legged platforms.
Paper Structure (19 sections, 6 equations, 7 figures, 2 tables)

This paper contains 19 sections, 6 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Creating a High-Quality, Real-World Dataset
  5. Mapping
  6. Synthetic pointclouds
  7. Traversability Prediction with UNRealNet
  8. Probabilistic Traversability Function
  9. Experiments, Results and Analysis
  10. Network Performance
  11. Does UNRealNet Outperform Standard Baselines?
  12. Does the Probabilistic Traversability Function Better Capture Ground-Truth?
  13. Hardware Experiments
  14. Robot-Agnosticism
  15. Kalman Filter
  16. ...and 4 more sections

Figures (7)

  • Figure 1: We propose UNRealNet (b), a network that produces high-quality, robot-agnostic navigation features and traversability (c) directly from pointclouds from an on-board sensor (a). Colormap for risk maps included for clarity.
  • Figure 2: An overview of our algorithm. We first create a high-quality map (1) from a pointcloud from a laser scanner. We then generate synthetic pointclouds (2) and their corresponding ground-truth map features (3). We train UNRealNet to predict the high-quality, robot-agnostic map features from the noisy, synthetic pointclouds (4, 5). We can then deploy this network on multiple robots (6) by leveraging a robot-specific traversability function (7). Red arrows denote network inputs, blue networks denote network outputs, and the orange arrow denotes the training objective (Equation \ref{['eq:loss_fn']}).
  • Figure 3: An example that highlights the value of the high-resolution scanner. (a) the pointcloud from the scanner. (b) the slope feature of the corresponding local map. We are able to identify the edges of each wire. (c) A photo of the environment and (d) the slope feature from the registered pointcloud. The noise from the lidar and SLAM is too high to see the wires.
  • Figure 4: A visualization of a datapoint in our dataset, showing (a) the depth image obtained from simulating lidar in the ground-truth pointcloud (blue=close, red=far), (b) the depth image obtained by running the original depth image through the noising pipeline, (c) the BEV projection of the original pointcloud, (d) the BEV projection of noised pointcloud, and corresponding crops from the ground-truth elevation map (e) and a sample traversability map (f).
  • Figure 5: Traversability predictions using the same network output on a sample in env6. (a) A crop from the FARO-scanned environment. There is a pile of wires in the middle. (b) The ground-truth elevation map. (c) The traversability map generated from parameters for a Spot. (d) The traversability map generated from parameters for an AlienGo. (e) A mosaic of traversability maps by varying the local slope (y axis) and robot slope (x axis) thresholds.
  • ...and 2 more figures