Learning Risk-aware Costmaps for Traversability in Challenging Environments.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Autonomous robots face challenges navigating unknown environments due to uncertainty from localization errors, sensor noise, and unpredictable robot-ground interactions.
• Traditional geometric methods for terrain analysis are often computationally expensive and prone to modeling errors, especially when dealing with uncertain traversability costs.

Purpose of the Study:

• To develop a robust learning-based method for estimating traversability costmaps in unknown environments.
• To address the challenge of uncertainty in robot navigation by learning tail-risks, specifically conditional value-at-risk (CVaR).

Main Methods:

• Introduced a neural network architecture designed to learn the distribution of traversability costs.
• Employed a tail-risk learning approach, focusing on conditional value-at-risk (CVaR), to manage uncertainty.
• Validated the method using data from a legged robot in diverse, challenging environments like subway tunnels and caves.

Main Results:

• The proposed method reliably learns expected tail risks for a given probability threshold.
• Generated traversability costmaps that are more robust to outliers and accurately capture tail risks.
• Demonstrated improved computational efficiency compared to existing baseline methods.

Conclusions:

• The learning-based tail-risk approach provides a more robust, accurate, and efficient solution for autonomous robot navigation in uncertain environments.
• This CVaR-focused method enhances robot safety by better managing the risks associated with unpredictable terrain conditions.