The field of robotics and imitation learning is moving towards developing more robust and generalizable models that can learn from sparse and noisy demonstrations. Researchers are exploring new approaches to address the challenges of demonstration noise and coverage limitations, and to enable effective learning from interacting with the environment. One of the key directions is the use of data augmentation techniques, such as generating potential transitions between demonstration skills and establishing unique connections for arbitrary states within the demonstration neighborhood. Another direction is the development of novel reward models that can estimate expert and behavioral distributions within the latent space of the world model, allowing for more stable and efficient learning. Furthermore, there is a growing interest in applying imitation learning to real-world humanoid robots, with a focus on whole-body control and teleoperation. Notable papers in this area include SkillMimic-V2, which presents a robust and generalizable approach to learning interaction skills from sparse and noisy demonstrations, and TWIST, which demonstrates a teleoperated whole-body imitation system for humanoid robots. Additionally, papers like PARC and ADD showcase innovative approaches to physics-based augmentation and adversarial multi-objective optimization for character controllers.