Introduction
Autonomous navigation systems often rely on classical stacks that require extensive manual tuning by experts. While robust, these systems typically use a static set of hyperparameters that fail to generalize across diverse environments. For instance, parameters optimized for high-speed traversal in open spaces may lead to collisions in narrow corridors. Conversely, end-to-end learning approaches, while flexible, often lack the safety guarantees of classical planners.
APPLD (Adaptive Planner Parameter Learning from Demonstration) bridges this gap by allowing non-experts to tune navigation stacks through simple teleoperated demonstrations. By treating the navigation stack as a "black box," APPLD learns context-specific parameters that dynamically adapt to the robot's surroundings.
Methodology
The APPLD framework consists of a four-stage pipeline designed to automate the parameter tuning process without requiring modifications to the underlying navigation algorithm.
- Demonstration & Segmentation: A human operator provides a single teleoperated run. The system employs Bayesian changepoint detection (CHAMP) to segment the trajectory into distinct environmental contexts based on LiDAR data and control inputs.
- Parameter Learning: For each segment, a black-box optimizer (CMA-ES) identifies the parameters that minimize the discrepancy between the autonomous planner's behavior and the human demonstration via Behavioral Cloning.
- Context Prediction: A neural network is trained to classify the current environment context using real-time LiDAR scans.
- Adaptive Deployment: During operation, the robot identifies its current context and swaps to the corresponding optimized parameter set on the fly.
Experimental Evaluation
The system was evaluated on two platforms: a Clearpath Jackal in a multi-zone maze and a BWIBot service robot in a real-world office hallway. In both cases, the training data consisted of a single demonstration lasting less than one minute.
Key Results
APPLD demonstrated significant improvements over default factory settings and even expert-tuned parameters.
- Efficiency and Reliability: In the Jackal "Curve" context, APPLD reduced traversal time from 30.20 seconds (default) to 9.39 seconds. In narrow corridors, where default parameters failed in 80% of trials, APPLD achieved a 100% success rate.
- Context Sensitivity: The results highlighted the "averaging trap"—optimizing a single parameter set for the entire environment (APPLD No-Context) led to suboptimal performance compared to the context-switching approach.
- Expert-Level Performance: On the BWIBot platform, APPLD achieved a lower Behavioral Cloning loss (0.0669) than parameters manually tuned by experts over several days (0.0940).
Discussion and Future Work
A key advantage of APPLD is its ability to preserve the safety guarantees of classical planners while incorporating the adaptability of learning-based methods. Interestingly, the system often outperformed the human demonstrator by eliminating reaction-time delays and providing smoother control.
Current limitations include the offline optimization time (approximately 8 hours) and the requirement that the human demonstrator encounter all relevant contexts. Future research aims to develop a "universal library" of parameters by clustering similar contexts across different environments, potentially enabling zero-shot navigation in novel buildings.
Summary updated