Why Do Humanoid Robots Still Struggle With the Small Stuff?

2026-03-13 · Source: artificial intelligence – Quanta Magazine · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Intermediate, long

Summary

Humanoid robotics has seen significant advancements since 2015, moving from frequently falling prototypes to more agile systems like Boston Dynamics' Atlas and Agility Robotics' Digit. These improvements stem from three paradigm shifts: deep learning enhancing computer vision and reinforcement learning, the adoption of proprioceptive electric motors replacing heavy hydraulics for nimbleness, and large language models enabling autonomous multi-step task planning. Despite these breakthroughs, including breakdancing robots and those handling irregular items, core challenges like reliably navigating stairs and opening doors remain. Experts attribute this to the incomplete mastery of physics, specifically force and inertia control, which is crucial for delicate manipulation and human-like interaction speeds. Current AI models often learn position-based control, with force regulation happening indirectly, leading to limitations in precision and safety when interacting with complex, dynamic environments.

Key takeaway

For AI Scientists and Research Scientists developing humanoid robots, recognize that while deep learning and advanced actuation have improved robot agility, explicit force control remains a critical, unsolved challenge. Your focus should shift towards integrating physics fundamentals, such as force and acceleration, directly into AI architectures and learning compliant behaviors in simulation to achieve true multipurpose mobile manipulation and safe interaction with delicate objects, rather than relying solely on position-based control.

Key insights

Despite AI advancements, humanoid robots still struggle with fundamental physics like force control for reliable real-world interaction.

Principles

• Proprioceptive actuators enable robust, compliant robot hardware.
• Deep reinforcement learning drives complex whole-body coordination.
• Multimodal AI unifies perception, planning, and control.

Method

Neural networks are trained as whole-body controllers using countless digital simulations, learning policies to translate environmental feedback into actions, often replacing hand-engineered algorithms.

In practice

• Utilize proprioceptive electric motors for compliant robot design.
• Employ deep reinforcement learning for dynamic locomotion.
• Integrate vision-language-action models for task planning.

Topics

• Humanoid Robotics
• Reinforcement Learning
• Force Control
• Proprioceptive Actuators
• Vision-Language-Action Models

Best for: AI Scientist, Research Scientist, AI Researcher, Robotics Engineer, AI Engineer

Related on AIssential

• This Humanoid Robot Is a Terrifyingly Competent Office Intern — Flexion Robotics trains humanoids by combining simulated skill learning with a master AI interpreting human video actions.
• PressMimic: Pressure-Guided Motion Capture and Control for Humanoid Robot Imitation — Pressure serves as a unified physical grounding signal, bridging perception and control for robust humanoid motion imitation.
• Humanoid Robots Exit Labs: Mapping the Technical Path to Embodied AI at AW 2026 — Humanoid robots are evolving into embodied AI platforms, transitioning from labs to industrial applications via data-driven iteration and distributed computing.
• Proprioceptive-visual correspondence enables self-other distinction in humanoid robots — Humanoid robots can learn self-other distinction via proprioceptive-visual correspondence, enabling a 3D self-model.
• Why Robots Still Struggle With Simple Tasks (And What Might Finally Change That) | Karol Hausman, Co-Founder & CEO of Physical Intelligence — Combining internet-scale LLM knowledge with robot motion data enables general physical intelligence for complex tasks.
• Why Touch Is Becoming a Big Deal Again in Robotics — Touch sensing is crucial for robots to understand physical interaction, enabling robust manipulation and adaptive control beyond vision.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by artificial intelligence – Quanta Magazine.