TL;DR
This paper introduces KHAOS, a novel 3D reactive planning system for drones operating near humans. KHAOS uses stochastic optimization to generate socially acceptable trajectories by considering drone kinematics, human discomfort, and visibility, enabling safer and more intuitive human-drone collaboration.
What Problem Does This Paper Address?
As drones become more integrated into human-populated areas for tasks like delivery and collaboration, their ability to navigate safely and interact naturally with people is crucial. Current navigation systems often prioritize efficiency or obstacle avoidance, neglecting the social cues and comfort of nearby humans. This can lead to unsettling or even frightening drone behavior, hindering widespread adoption. The paper specifically addresses the challenge of designing drone trajectories that are not only kinematically feasible and safe but also socially aware, minimizing human discomfort and maximizing visibility, especially in scenarios like "flying co-worker" applications where direct interaction is expected.
Methods
The core of KHAOS is a reactive trajectory planner built upon the STOMP (Stochastic Trajectory Optimization for Motion Planning) algorithm. This approach uses stochastic optimization to explore a space of possible trajectories. KHAOS extends STOMP by incorporating several key components:
- Kinematic Constraints: The system adheres to the drone's physical limitations, including maximum speed, acceleration, and deceleration.
- Human-Aware Costs: Two novel cost functions are introduced to guide the optimization towards socially acceptable trajectories:
- Discomfort Cost ($C_{dis}$): This cost quantifies the unease caused to humans by a drone's proximity and speed. It considers both the relative speed and distance between the drone and humans, penalizing fast movements close to people and encouraging speed reduction or stopping when necessary.
- Visibility Cost ($C_{vis}$): This cost promotes the drone's visibility to humans. It's modeled as a 3D grid representing the human visual field, assigning lower costs to areas directly in front of the human and higher costs to peripheral or hidden areas.
- Constrained Optimization: The optimization process balances these costs with kinematic constraints. A "discomfort constraint" ($DCF_{max}$) is introduced, which limits the drone's speed if the calculated discomfort exceeds a predefined threshold, even if kinematically possible. This ensures that the drone prioritizes human comfort over maximum speed when necessary.
- Reactive Planning: The system is designed to be reactive, meaning it can re-plan trajectories on the fly in response to changes in the environment or human positions.
Datasets
Not discussed in the paper. The evaluation is performed in simulated environments using the MoveIt! collision scene.
Key Insights and Ideas
- Quantifying "Socially Acceptable": The paper moves beyond basic proxemics by proposing concrete, quantifiable metrics for human discomfort and visibility, allowing for their integration into optimization-based planning.
- Discomfort as a Speed Regulator: The discomfort cost acts as a powerful mechanism to naturally regulate drone speed, making it slow down when near humans, which is intuitively more comfortable and less startling.
- Visibility for Intent: The visibility cost encourages trajectories that keep the drone within a human's field of view, which is crucial for clear communication of intent and for collaborative tasks.
- Balancing Kinematics and Social Cues: KHAOS demonstrates that it's possible to effectively balance the physical limitations of a drone with complex, human-centric social considerations within a unified optimization framework.
- Adaptability to Constraints: The system shows a strong ability to adapt its trajectory not just locally but by exploring different "homotopy classes" (path shapes) to find more suitable and comfortable solutions, even in highly constrained environments.
Key Results
The paper presents qualitative results from various simulated scenarios, demonstrating KHAOS's capabilities:
- Frontal Approach: The drone successfully approaches a human, adjusting its speed to stay within the discomfort constraint, showing smoother deceleration phases when the constraint is lowered.
- Two-Human Scenario: KHAOS navigates a corridor with two humans, prioritizing visibility and minimizing discomfort for both, even when faced with wall constraints. The planner effectively manages the combined influence of multiple humans.
- Planner Reactivity: In a corridor scenario, when a human's position is shifted, the planner immediately reacts by deviating the path and reducing speed, showcasing its real-time adaptability.
- Highly Constrained Environment: In a challenging corridor with an added obstacle, KHAOS finds a more suitable trajectory that avoids passing too close to the human, demonstrating its ability to explore alternative paths rather than just locally deforming the initial one.
The planner operates at a frequency of 5-10Hz on a standard computer, suggesting its potential for real-time application.
Limitations and Future Work
- Drone Orientation: The current implementation does not explicitly consider the drone's orientation as a social constraint, which could be important for human perception.
- Smoothing: While the system produces smooth motions, further improvements could be made by incorporating bounded jerk models for even better velocity profiles.
- Handover Phase: The paper suggests extending the planner to manage the handover phase of an object, which would require closer proximity and coordination with a manipulator arm.
- Deformable Configuration Spaces: Testing deformable configuration spaces for the drone is proposed for more complex interactions.
- User Studies: The authors plan to conduct user studies to tune social constraint parameters and assess the social acceptability of KHAOS-generated trajectories, gathering feedback for real-world robot experiments.
Summary updated