The Scientific Challenges to Making Humanoid Robots Smarter (AI Translation)

【周刊提前读】要让人形机器人更智能 还要攻克哪些科学难题？

[para. 1] In January 2025, Jacob Delesio received a humanoid robot G1 from Hangzhou Unitree Robotics Co. Delesio unboxed the robot, which demonstrated agility and performed a "dynamic reset," a maneuver involving biomechanical properties and algorithms. [para. 2] Later that month, 16 humanoid robots from Unitree Robotics showcased their skills at the Spring Festival Gala, performing in coordination with human dancers. The performance highlighted advances in dynamic motion and coordination of robots. [para. 4] This is largely due to the integration of AI technologies, particularly large language models, which have improved robots’ planning and execution capabilities.

[para. 5][para. 8] The evolution of humanoid robots mimics human evolutionary processes, where motion control, sensory information processing, and task execution play crucial roles. Advancements in these areas allow robots to better interact with their environment. [para. 9][para. 10] The journey of robotics began with rudimentary tasks performed by early industrial robots, evolving to complex operations driven by motion control technology. The introduction of deep learning enabled significant improvements in robotic sensory and motion capabilities. [para. 12][para. 13][para. 14] Today's robots benefit from neural networks, which provide robust learning and recognition capabilities, a stark contrast to the rigid program logic of early machinery.

[para. 22][para. 23] Motion control has evolved over time, with groundbreaking implementations like the SCARA robot for precise positioning tasks, to Boston Dynamics' BigDog, which marked a step towards robotic autonomy via its sensor data processing. [para. 26][para. 29] Technologies from AI research fields like deep neural networks and reinforcement learning introduced with AlphaGo have dramatically advanced robots’ ability to plan and execute complex actions autonomously. The integration of these technologies in robotics fosters adaptive capabilities similar to human learning processes.

[para. 34][para. 35][para. 36] Recent developments in embodied intelligence suggest advances in robots' real-world interactions, particularly in understanding and executing tasks in domestic environments. Experiments with robotic arms at Tsinghua University's Embodied Intelligence Lab highlight challenges in spatial generalization, where the robots adapt to varied object placements with precision. [para. 37][para. 39] Despite progress, current algorithms still struggle with visual bias and environmental adaptability, pinpointing areas for future improvements in generalization and perceptual accuracy.

[para. 45][para. 46] Wang Yu emphasized improvements in system learning capacities through self-learning models, which enable tasks previously unattainable for robots. [para. 48][para. 50] While generalized robotic intelligence remains a difficult goal, projects like OpenAI’s Figure 01 highlight strides in robotic adaptability and environmental interaction, laying groundwork for humanoid robots with human-like dexterity. [para. 51][para. 55] Yet, challenges persist, primarily in developing finely tuned mechanical systems and dexterous hands that can match human capabilities.

[para. 61][para. 62] Current limitations in robotic tactile sensing and multi-degree-of-freedom movements underscore inherent physical and algorithmic challenges. Solutions require breakthroughs in sensor technology and deeper understanding of motion dynamics within real-world contexts. [para. 73][para. 75] Beyond physical challenges, artificial intelligence research is guided by principles akin to traditional scientific exploration, where rigorous understanding and precise modeling are key to achieving transformative advances. While robots might mimic human actions, achieving genuine machine intelligence necessitates ongoing inquiry into the fundamental nature of intelligence.