Humanoid robots have advanced dramatically over the past decade, yet many still struggle with tasks that humans consider simple, such as climbing unfamiliar stairs, opening doors, or manipulating fragile objects. Despite impressive demonstrations from robots such as Boston Dynamics’ Atlas and Agility Robotics’ Digit, researchers say these systems are far from mastering the everyday physical interactions required for widespread deployment, tells Quanta Magazine.
Several technological breakthroughs have pushed humanoid robotics forward. Deep learning and reinforcement learning have enabled robots to develop coordinated motion by training neural networks in simulated environments. These models learn “policies” that translate sensory input into movement commands, allowing robots to maintain balance, avoid collisions, and navigate complex spaces. At the same time, new electric actuators with built-in flexibility have replaced older hydraulic systems, giving robots more agile and animal-like motion. More recently, large language models and vision-language-action systems have helped robots interpret instructions and plan multistep actions.
Yet these advances have not fully solved the core challenge of robotics: controlling physical forces. Human actions depend on a deep understanding of force, pressure, and inertia acquired through years of interacting with the environment. Robots, by contrast, often rely on position-based control, which focuses on moving limbs to specific locations without fully sensing the forces involved. This approach works for simple movements but breaks down when robots interact with unpredictable objects or delicate materials.
Without precise force control, robots must operate slowly or cautiously to avoid damaging objects or themselves. Their rigid bodies and high inertia further complicate delicate manipulation, making it difficult to replicate the nuanced touch humans use when performing routine tasks such as twisting a bottle cap or handling an egg.
Researchers believe future progress will require combining improved hardware, such as tactile sensors and more compliant robotic hands, with new learning algorithms that explicitly incorporate the physics of force and contact. Some scientists advocate large-scale training datasets and pretrained behavior models, while others argue for fundamentally new AI architectures grounded in physical principles.
For now, humanoid robots remain powerful but incomplete machines. As one roboticist put it, the field may still be in an early experimental phase, where impressive demonstrations exist but the deeper scientific understanding needed for reliable, everyday autonomy is still emerging.