Risk-aware path planning method for mobile robots in planetary environments (ICRA 2023)

We are excited to present our new work on risk-aware path planning for planetary exploration robots (rovers) at the IEEE International Conference on Robotics and Automation (ICRA 2023)!

Masafumi Endo, Tatsunori Taniai, Ryo Yonetani, and Genya Ishigami, “Risk-aware Path Planning via Probabilistic Fusion of Traversability Prediction for Planetary Rovers on Heterogeneous Terrains,” IEEE ICRA, 2023. [arXiv] [Project] [Code].

Welcome

ICRA has been recognized as one of the prestigious conferences for robotics and automation technologies. This work was achieved during my internship at OMRON SINIC X, in collaboration with my advisor, Prof. Genya Ishigami, from Keio University.

Motivation

This work addresses path planning for wheeled rovers exploring distant planets. Path planning is an essential component of autonomous navigation on celestial surfaces where communication latency disturbs manual rover operation (e.g., 5 to 20 minutes between Earth and Mars).

Typical rover, such as NASA’s Perseverance, is based on wheel locomotion. Courtesy NASA/JPL-Caltech.

Unlike typical path planning problems for well-structured on-road environments, autonomous navigation in such remote off-road environments requires careful assessment of the degree of traverse ability, or traversability. Wheel slip on deformable surfaces is a prominent factor of traversability degradation, impeding driving speed and eventually causing permanent rover entrapment.

During the MER mission 2005, the Opportunity rover experienced a significant wheel slip on sandy terrain. Courtesy NASA/JPL-Caltech.

When the wheel slip is considered as the traversability, it can be quantified as the slip ratio s, based on the comparison of actual velocity u and reference velocity u_ref, as follows:

A positive slip ratio represents a traverse slower than commanded, with s = 1 representing complete rover entrapment in deformable terrains. Conversely, a negative slip ratio expresses a traverse faster than commanded.

We use this slip ratio to construct cost maps for path planning on planetary surfaces. But how can we predict slip on those surfaces if there are various slip-inducing terrains, such as pebbles, loose sand, and fine gain?

Planetary surface conditions are covered with various heterogeneous terrains. Courtesy NASA/JPL-Caltech.

Ultimately, we could investigate detailed material compositions to find their correlation with slip tendency, but it’s not feasible for environments yet to be explored. Visual information of terrain can be exploited to assess such slip tendency of unexplored terrain. Machine learning (ML) algorithms are vital in such vision-based slip prediction. Suppose we have environment information based on pixel-wise appearance and geometry (3D position) information and plan paths in the graph associated with visual information. ML models are commonly used for the following objectives.

• Symbolically categorize given appearance information into distinctive terrain classes using ML classifiers (semantic segmentation models).
• Model latent slip (LS) functions that cast terrain geometry (e.g., terrain pitch angle) to slip ratio for identified terrain classes using ML regressors.

Typical classify-then-regress-based pipelines for slip prediction from visual information.

If the ML predictions are perfect, we could successfully plan paths using cost maps constructed from the predictions. However, learning-based approaches inevitably suffer from potential prediction errors. Such potential errors come from imperfect ML models and limited observations on distant planets. As a result, erroneous traversability prediction can cause unrecoverable rover immobilization.

We here argue the central question in this work:

How to address inevitably erroneous traversability prediction to construct cost maps in path planning?

Our Solution

This work presents a novel path-planning method that explicitly accounts for erroneous traversability prediction from ML models. Our key idea is the probabilistic fusion of distinctive ML models while considering their uncertainties. We employ two pre-trained ML models:

• Convolutional neural network (CNN) to pixel-wisely predict terrain classes from appearance imagery.
• Gaussian processes (GPs) to model class-dependent LS functions with geometry information.

For every graph vertex, we first predict a categorical distribution of terrain classes and a slip distribution for corresponding terrains with a given slope. We then integrate multiple slip distributions weighted by the classification likelihoods to form a single distribution as mixtures of GPs (MGP).

Multiple probabilistic distributions for terrain-dependent traversability at arbitrary positions (top) are fused into a single, multimodal slip distribution via MGP (bottom).

This expression gives us a multimodal slip distribution accounting for heterogeneous terrains and retains the prediction uncertainties in both ML models. The fused distribution further enables us to evaluate less-likely occurring slips in the MGP distribution via conditional value at risk (CVaR).

CVaR offers conservative risk estimates of slip from the slip distribution given by MGP.

By translating predicted slip to velocity, we can estimate a travel time cost that considers both traversability in terms of risk and traverse efficiency.

Simulation Experiments

We have extensively investigated our proposed method in simulated environments with varying difficulties in traversability prediction. Here, we prepare three synthetic datasets well-replicating situations that rovers will encounter on celestial surfaces, as

• Standard (Std): This dataset adopts a simple setting where pixel color identifies terrain classes and slip measurements contain slight noise.
• Erroneous Slip (ES): A more complicated, realistic setting is considered in this dataset by imposing greater noise in slip measurements. Hence, learned GPs result in erroneous slip models.
• Ambiguous Appearance (AA): We replicate another complicated setting where multiple terrain classes share the same appearance. This can happen when distinctive terrain characteristics do not provide visual cues or another covers one terrain material.

From left to right, input maps (height and color), color maps, ground truth terrain class maps, and GPs showing learned slip functions.

As baselines for comparison, we implemented several variants of the proposed method (denoted as MGP+CVaR). SGP methods use a single GP selected from terrain classification results [C. Cunningham et al., IEEE ICRA, 2017] instead of using MGP. EV methods use the expected value for probabilistic slip estimation instead of CVaR, and VaR methods use the value at risk.

Qualitative comparisons illustrate that our method based on MGP and CVaR successfully navigates the rover on a less-slippery path than the existing approaches, even in erroneous ML predictions.

Visualized planning results over different datasets.

Quantitative path planning results further validate that our proposed approach has the highest safety performance throughout the datasets, as shown by its success rates (Succ.), indicating path execution rates without immobilization, and maximum slips (s_max), indicating the maximum slip ratio during path execution. Our method enables safe and efficient rover navigation for the following reasons.

• The MGP-based slip distribution modeling provides greater expressive power than a single GP, effectively accounting for heterogeneous terrain as well as overall uncertainties from classification and regression models.
• CVaR-based risk assessment provides rationally conservative slip estimates by looking at rare but risky slip occurrences in the distribution.
• Having lower slip contributes to time efficiency by retaining rover velocity.

Quantitative path planning results over different datasets.

Summary

• We presented a novel path planner that accounts for uncertainty in ML-based traversability prediction on heterogeneous deformable terrains.
• Our probabilistic pipeline enables multimodal slip prediction while incorporating the uncertainty from different ML models into the costs of path planning.
• Our method provides safer paths under visual ambiguity and unreliable slip models, outperforming existing classify-then-regress methods.

Learn More About Our Work

• Our paper explains more details of our work.
• Our project site also summarizes useful material, such as the code repository, video presentation, etc.
• The authors are welcome to questions and discussions in-person at ICRA or via code repository.