QDM-RNN: Acquisition of High-Speed and Robust Behavior from Low-Speed Demonstrations

Imitation learning has gained significant attention as a promising approach to enable flexible and generalizable robot motion generation across diverse tasks. However, existing models often suffer from long inference times, limiting their applicability to high-speed and fine-grained tasks. Moreover, while faster computation enables shorter inference cycles, it introduces new challenges such as overcontrol and motion instability when the inference frequency exceeds the sensor sampling rate. To address these issues, we propose QDM-RNN, a lightweight motion generation model that learns from slow but high-quality demonstrations and remains robust under high- frequency inference. Our method utilizes Softmax Transfor- mation, which discretizes the robot end-effector pose into a high-resolution probability distribution, enabling accurate and smooth trajectory prediction. Furthermore, by incorporating multi-timestep prediction, which simultaneously predicts sev- eral future steps, our model mitigates instability arising from the mismatch between sensor and inference rates, ensuring consistent long-horizon motion generation. We validated the effectiveness of our approach through real-world robotic ex- periments on a Sport Stacking task, which requires both high speed and precision. QDM-RNN maintained high success rates and motion stability at speeds up to seven times faster than demonstrations, and showed robustness to external disturbances including background changes, lighting variations, dynamic object perturbations, and obstacles.