Humanoid robotics has made significant strides in recent years, driven by advancements in deep learning, electric actuators, and large language models. Despite this progress, experts warn that humanoids still struggle with fundamental physical tasks like opening doors or navigating stairs. In 2015, bipedal robots were prone to falling, but today's machines, such as Boston Dynamics' Atlas and Agility Robotics' Digit, move with surprising grace. However, both Scott Kuindersma of Boston Dynamics and Jonathan Hurst of Agility Robotics admit these systems do not reliably handle diverse staircases or doorways yet. The field has evolved through three major technological shifts. First, deep learning on powerful GPUs has improved computer vision and reinforcement learning, enabling robots to perceive environments faster. Second, the replacement of heavy hydraulics with lightweight, compliant electric motors allows for nimble, animal-like movement. Third, large language models adapted for robotics facilitate multi-step task planning. These innovations have transformed rigid, halting machines into fluid performers capable of complex actions in controlled demos. Yet, a critical gap remains: the mastery of force and physics. While current AI excels at positional control, it lacks the intuitive understanding of force and inertia that humans possess through their musculoskeletal systems. Pulkit Agrawal of MIT notes that to achieve human-like dexterity, robots must explicitly control forces. Classical force control, which treats robot joints as springs and dampers, is effective for specific industrial tasks but difficult to generalize. Modern robotics often relies on neural networks that learn positions in simulation, indirectly approximating force regulation. This limitation becomes apparent when robots interact with delicate objects or navigate complex environments. Without direct force sensing or control, stiff, heavy humanoid bodies risk damage or failure during precision tasks. As Agrawal explains, robots currently compensate for this by moving slowly, a strategy that sacrifices efficiency. Carolina Parada of Google DeepMind acknowledges that vision-language-action models can generate movement steps, but they often rely on defined poses rather than understanding the physical resistance of the environment. Experts debate how to bridge this gap. Some, like Agrawal, advocate for combining reinforcement learning with explicit force control in simulations. Others, such as MIT's Russ Tedrake, suggest that current hardware is sufficient and the focus should be on better control algorithms and large-scale data collection. Frank Park from the University of Southern California goes further, arguing that current AI architectures may be fundamentally flawed and that foundational physics principles must be integrated into AI learning from the ground up. The scientific community views this phase as a transitional period similar to the early days of electricity research, where practical applications preceded a complete theoretical understanding. While the hardware is robust and the potential for general-purpose dexterity is clear, the industry recognizes that true autonomy requires more than just better AI or sensors. As Tedrake concludes, the bones of the technology are sound, but the field still has much to learn. Humanoid robots remain promising, yet solving the physics of movement is an ongoing challenge that will take time to overcome.