In this project, we attempt to find out whether RL methods can be useful for 3D lifting of the human pose from 2D data. This task is very important in biomechanics, as it may contribute to inclusivity in the field of study as well as clinical applications, as not all labs or clinics can afford state-of-the-art motion capture systems. There are several existing 3D-lifting methods developed (e.g., MotionBERT, OpenCap). These methods are also trained to optimize for reconstruction loss, which gives them an opportunity to get over confident about the errors they make, and the models are typically quite large. More light-weight lifters exist, but naturally under-perform bigger models. We will look into training a small correction network using RL methods to bridge the gap to the ground truth, training on openly available labeled data (mocap + RGB-video) from AMASS. This would give us a more controlled setting to build the RL training setup in a more predictable way before generalizing to a more complex motion. We want to train a small correction network that takes 3D pose estimates produced by a 3D lifting model (e.g. MotionBERT) as input and nudges them toward more physically plausible configurations - without using any additional 3D labels at correction time. The correction network is trained using reinforcement learning. Its reward signal comes from two sources: a consistency check against the original 2D observations (geometric grounding), and a learned GAIL-like discriminator that has been trained to distinguish real human motion from MotionBERT estimates (motion naturalness). The discriminator reward is the inverse reinforcement learning component, it encodes what good motion looks like by learning from expert demonstrations in the AMASS mocap dataset, rather than from any hand-designed rule. More physics-derived terms could be added to the reward to further ensure naturally looking motions, or frame-to-frame smoothness, if the discriminator score should prove insufficient.