TLDR: A new drone trajectory planner, based on Decision Transformers (DT), allows dynamic adjustment of safety and efficiency using a single parameter called Return-to-Go (RTG). Unlike traditional methods requiring complex tuning or retraining, this approach enables a drone to prioritize safety (lower RTG) or efficiency (higher RTG) intuitively. Validated in simulations and real-world tests, the planner outperforms existing methods and offers flexible, real-time control over drone behavior in diverse environments.
Drone technology continues to advance rapidly, finding applications in diverse fields from delivery services to environmental monitoring. A critical challenge for autonomous drones is planning trajectories that effectively balance safety and efficiency, especially in unpredictable environments. Traditional methods often fall short: polynomial-based planners, while computationally efficient, demand extensive expert knowledge to fine-tune parameters for this trade-off and tend to prioritize efficiency. Reinforcement learning (RL)-based planners, though adaptable, don’t explicitly address the safety-efficiency balance and typically require costly retraining for different mission requirements.
A Novel Approach with Decision Transformers
To overcome these limitations, researchers have introduced a groundbreaking drone trajectory planner based on the Decision Transformer (DT) framework. This innovative approach leverages a single, intuitive parameter called Return-to-Go (RTG) to dynamically adjust the safety–efficiency trade-off. Unlike previous methods, this DT-based planner eliminates the need for expert knowledge in parameter tuning, separate algorithms, or extensive retraining processes.
The core idea is that RTG acts as a ‘temperature parameter’. In this framework, a higher RTG value instructs the drone to prioritize efficiency, leading to trajectories that might fly closer to obstacles to reach the target faster. Conversely, a lower RTG encourages the drone to prioritize safety, generating trajectories with larger safety margins and avoiding obstacles more conservatively. This means a single trained DT model can produce a wide range of behaviors simply by adjusting this one input parameter.
Enhancing Reliability with N-step RTG
A common challenge with conventional DT methods is the uncertainty introduced by RTG, which traditionally represents cumulative rewards over an entire episode. To make RTG more effective and reliable in unknown environments, the researchers introduced an ‘N-step RTG’. This limits the prediction horizon to a short, predictable window of N future steps, allowing for more consistent and effective behavior adjustments during flight.
Also Read:
- VLMPlanner: Enhancing Autonomous Driving with Visual Language Models and Adaptive Reasoning
- Enhancing Robot Planning Through Observation-Based Learning
Training and Validation
The DT-based planner was trained using a diverse dataset of expert drone trajectories. This data was collected from state-of-the-art planning algorithms, including EGO-Planner and ViGO, known for their efficiency and collision avoidance. Crucially, the dataset also incorporated trajectories from a newly devised medial axis-based planner, which explicitly prioritizes safety by maximizing clearance from obstacles. This diverse training data enabled the DT model to learn both safe and efficient flight behaviors.
Extensive simulations were conducted in Gazebo, using both structured grid maps and unstructured random environments, including scenarios not seen during training. The results consistently demonstrated that the proposed planner could dynamically adjust the safety–efficiency trade-off by simply tuning the RTG parameter. When tuned for safety (lower RTG), the planner generated trajectories with superior safety performance. When tuned for efficiency (higher RTG), it achieved better flight times and average velocities compared to existing baseline methods.
Furthermore, the ability to adjust RTG in real-time during trajectory generation was validated, showing that the drone could seamlessly switch between safe and efficient behaviors mid-flight. Real-world experiments using a physical drone further confirmed the reliability and practicality of the proposed algorithm, mirroring the simulation findings and proving its effectiveness in actual drone operations.
This research marks a significant step forward in autonomous drone navigation, offering a flexible and robust solution for dynamic trajectory planning. For more details, you can refer to the full research paper here.
