Have you looked / interacted with Pepper and Nao robots? In my experience, Aldebaran Robotics tried hard to go down the "lifelike robot" route, and the issues they encountered are real and hard problems.
A big problem with body control was lack of touch / impact feedback. Do you have plans to integrate sensors on the robot's 'skin'?
Modern ML based perception is expensive (audio, visual, NLP, etc). Modern ML-based full-body planning even more so. Do you intend to go down a hand-crafted route for walking, gestures, grasping, or do you intend to do it all with ML?