Learning Human Rhythmic Movements: Adaptive CPGs for Synchronized Virtual Agents
Résumé
This paper presents a real-time framework for enabling a virtual agent to learn and synchronize with rhythmic human gestures using bio-inspired control. Each joint-axis of the agent is controlled by a Central Pattern Generator (CPG), a small network of adaptive Hopf oscillators, which continuously learns amplitude, frequency, phase, and offset from streaming motion-capture input. The pipeline maps Kinect V2 skeleton data to a normalized avatar representation, computes per-frame error signals, and updates 75 parallel CPGs at 30\,Hz, allowing fast convergence and real-time operation. Evaluated on a set of ten 60-second rhythmic gestures performed by a professional dancer, the system demonstrates that the CPGs quickly adapt to the human motion and can autonomously reproduce the learned trajectories after the learning process is frozen. This approach provides a compact and interpretable mechanism for sensory–motor coupling between human and virtual agent, with potential applications in creative co-dancing, interactive systems, and motor rehabilitation.