Abstract
Generating natural and synchronized robot dance choreography remains a significant challenge in robotics. Traditional methods often rely on fixed sequences or probabilistic models that may produce physically awkward transitions. This paper presents Plan2Dance, a system that treats choreography as an automated planning problem. By utilizing the Planning Domain Definition Language (PDDL), Plan2Dance ensures that generated dance moves are logically consistent, physically feasible, and temporally aligned with music. The system was validated using a library of 58 expert-defined motions and successfully executed on a humanoid robot.
Introduction
Creating a robot that dances naturally requires balancing variety with physical constraints. Previous approaches generally fall into two categories: pre-defined sequences, which lack adaptability, and probabilistic models like Markov Chains, which can lead to "unreasonable" transitions due to a lack of hard logical constraints. Plan2Dance addresses these issues by framing choreography as a formal planning task, ensuring that every move follows the logical and temporal rules of dance.
Methods
The Plan2Dance pipeline is divided into three primary stages: music analysis, domain modeling, and automated planning.
- Music Analysis: Raw audio is processed using the
pyAudioAnalysislibrary to extract tempo and beat information. This data is used to automatically generate a PDDL problem file, which specifies the robot's initial state, the target dance duration, and preferences based on the music's characteristics. - Domain Modeling: In collaboration with dance experts, the authors developed a domain file containing 58 distinct action models. These models are defined in PDDL 2.1, allowing for the inclusion of temporal constraints (duration) and numeric features (preconditions and effects).
- Automated Planning: The system employs the OPTIC-CLP temporal planner. It processes the music-derived problem file and the expert-defined domain file to produce an optimal sequence of timed actions, which are then executed by a ROBOTIS MINI humanoid robot.
Choreography as Logic
The core innovation of Plan2Dance is the formalization of dance as a symbolic planning problem. By defining preconditions (e.g., the robot must be in a specific stance) and effects (e.g., the resulting posture), the system inherently prevents impossible transitions.
Furthermore, the use of temporal PDDL allows the "beat" of the music to serve as a hard constraint. This ensures that specific actions, such as a hand wave, occur precisely at the correct rhythmic interval. Preference-based planning is also utilized to match the dance's intensity to the music's speed, selecting between "fast" or "slow" variants of motions to optimize the overall aesthetic quality.
Results and Discussion
The system was evaluated using a library of 10 modern songs and 58 basic dancing motions. Plan2Dance successfully generated valid PDDL problems for all inputs and produced executable plans. Qualitative testing on the ROBOTIS MINI hardware confirmed that the generated choreography maintained high synchronization with the input audio while adhering to physical constraints.
Despite its success, the system currently relies on manual domain engineering, which presents a "knowledge bottleneck." Future work aims to automate the creation of action models by using machine learning to extract motion logic from videos of human dancers. Additionally, larger-scale user studies are needed to quantitatively compare the perceived quality of planning-based choreography against probabilistic baselines.
Conclusion
Plan2Dance demonstrates that automated planning is a robust framework for robot choreography. By combining music analysis with formal logic, the system bridges the gap between rhythmic synchronization and physical feasibility, providing a structured approach to creative robot motion.
Summary updated