Core Concepts
MotionRL is a novel approach that leverages reinforcement learning to fine-tune text-to-motion generation models, aligning them with human preferences and improving the quality of generated motions beyond traditional metrics.
Stats
MotionRL outperforms baseline models T2M-GPT and InstructMotion in R-Precision and FID scores on the HumanML3D dataset.
MotionRL achieves higher perceptual scores compared to other models based on the motion perception model from Wang et al. (2024).
Quotes
"almost all mainstream research has largely ignored the role of human perception in evaluating generated motions."
"generating realistic human motion, including smooth and natural movement, is more important than fitting existing error-based metrics, such as FID and R-Precision"
"Since such artifacts are difficult to measure using existing metrics (Zhang et al., 2023b), human perception of generated motions becomes crucial."