Abstract
We present a novel sensor-based learning navigation algorithm to compute a collision-free trajectory for a robot in dense and dynamic environments with moving obstacles or targets. Our approach uses deep reinforcement learning-based expert policy that is trained using a sim2real paradigm. In order to increase the reliability and handle the failure cases of the expert policy, we combine with a policy extraction technique to transform the resulting policy into a decision tree format. The resulting decision tree has properties which we use to analyze and modify the policy and improve performance on navigation metrics including smoothness, frequency of oscillation, frequency of immobilization, and obstruction of target. We are able to modify the policy to address these imperfections without retraining, combining the learning power of deep learning with the control of domain-specific algorithms. We highlight the benefits of our algorithm in simulated environments and navigating a Clearpath Jackal robot among moving pedestrians.
Introduction
Deep Reinforcement Learning (DRL) has become a cornerstone of modern robot navigation due to its ability to handle complex sensor data and dynamic environments. However, DRL policies are notoriously opaque "black boxes." When a robot exhibits undesirable behaviors—such as "freezing" in front of obstacles or oscillating wildly between paths—developers typically have few options beyond tweaking reward functions and initiating hours of expensive retraining.
The XAI-N (Explainable AI for Navigation) framework addresses this lack of interpretability and malleability. By distilling complex neural network policies into human-readable decision trees, researchers can identify specific failure modes and apply surgical logic fixes directly to the policy, bypassing the need for retraining while maintaining the performance benefits of deep learning.
Methodology
The XAI-N framework follows a structured three-stage pipeline to transition from a high-performance neural expert to an interpretable, refined controller.
1. Expert Policy Learning
The process begins by training a robust navigation policy using Proximal Policy Optimization (PPO). This expert policy is trained in simulation to navigate dynamic environments based on a 213-dimensional state space, which includes a radial occupancy grid derived from LiDAR data, goal coordinates, and previous actions.
2. Policy Extraction
To gain insight into the expert's decision-making, the neural network is distilled into a Decision Tree (DT) using the VIPER algorithm. This transformation converts the opaque weights of the neural network into a series of "if-then" statements based on sensor inputs.
As illustrated above, the decision tree partitions the robot's state space into discrete regions, each associated with a specific action. This structure allows researchers to pinpoint exactly which sensor configurations trigger problematic decisions.
3. Policy Modification
The final stage involves analyzing the extracted tree to identify and fix specific failure modes:
- Fixing Freezing: Nodes that command a "stop" action when obstacles are stationary are replaced with rotation commands to help the robot find an exit.
- Fixing Oscillation: Nodes causing rapid switching between left and right turns are split to introduce smoother, lower-velocity actions.
- Fixing Blocking: The tree is modified to ensure the robot does not stand in a human's path, improving social compliance.
Experimental Evaluation and Results
The researchers validated XAI-N across multiple environments, including Gazebo simulations and real-world deployments on a Clearpath Jackal robot navigating among pedestrians. The results demonstrate that the modified decision trees often outperform the original neural expert:
- Oscillation Reduction: In scenarios where the expert policy oscillated in 100% of runs, the XAI-N fix reduced this to just 6%, with the average oscillation length dropping from 8.07m to 1.33m.
- Eliminating Freezing: A policy that froze in 28% of scenarios was corrected to a 0% freeze rate.
- Efficiency Gains: The final policy achieved a higher average reward per timestep (0.241) compared to the original PPO expert (0.226), largely by spending less time in stalled or oscillatory states.
Conclusion
XAI-N demonstrates that the interpretability of decision trees can be leveraged to refine deep learning policies without retraining. While there are trade-offs—such as increased path length when prioritizing social compliance—the ability to surgically fix "stupid" behaviors provides a significant advantage for deploying reliable autonomous systems in the real world.
Summary updated