Science Friday

How AI Advances Are Improving Humanoid Robots

September 25, 2025

Key Takeaways Copied to clipboard!

  • While AI, particularly large language models like ChatGPT, significantly enhances a humanoid robot's conversational abilities, the more challenging aspects of robotics lie in achieving reliable locomotion and manipulation. 
  • The development of humanoid robots is being accelerated by advances in hardware, including 3D printing and the availability of commercial motors, making them more accessible for hobbyists and researchers. 
  • A major limitation in training robots is the vast data deficiency compared to large language models, and a critical area for future development is enabling robots to understand the 'why' behind their failures, similar to human self-awareness. 

Segments

Defining Humanoids and Toddy
Copied to clipboard!
(00:01:17)
  • Key Takeaway: Humanoid robots are generally defined by human-like morphology, including bipedal locomotion, manipulative hands, and egocentric vision systems.
  • Summary: A humanoid robot is characterized by a body structure similar to humans, featuring two legs for movement, two hands for interaction, and vision systems that allow it to perceive and navigate its environment from its own perspective. The Stanford-built ‘Toddy’ robot, modeled after a toddler, exemplifies these features, capable of speech, listening, and visual perception.
AI and Hardware Advances
Copied to clipboard!
(00:03:25)
  • Key Takeaway: Advances in both AI (like large language models) and robotics hardware (like 3D printing) are crucial for improving humanoid robots.
  • Summary: The progress in humanoid robotics is driven by two main factors: artificial intelligence, particularly large language models that enhance a robot’s reasoning and interaction capabilities, and significant improvements in hardware. Technologies like 3D printing and the use of commercially available motors have made building sophisticated robots more feasible and accessible.
Challenges in Robotics: Motion and Manipulation
Copied to clipboard!
(00:05:06)
  • Key Takeaway: Achieving reliable bipedal locomotion and general-purpose manipulation are the most difficult challenges in humanoid robotics, far exceeding the complexity of AI-driven conversation.
  • Summary: While conversational AI is becoming increasingly sophisticated, the fundamental physical tasks of walking and manipulating objects remain significant hurdles for robots. Bipedal locomotion is complex due to the under-actuated nature of the system, requiring precise coordination of joint torques to maintain balance and movement. Similarly, manipulation is challenging because robots need to interact with a wide variety of objects and environments with the same hardware and control policies.
Data Limitations and Learning Methods
Copied to clipboard!
(00:07:07)
  • Key Takeaway: Robots face a substantial data deficiency compared to AI models, necessitating diverse learning approaches like imitation and reinforcement learning.
  • Summary: Training robots requires significantly more data than training large language models, with current robotics datasets being orders of magnitude smaller. Robots can learn through imitation learning, supervised learning, and reinforcement learning, where they learn from trial and error by receiving scores or rewards for their actions. Designing effective reward functions that encompass ethical considerations is a complex task.
Ethical Considerations and Robot Intelligence
Copied to clipboard!
(00:10:07)
  • Key Takeaway: Ensuring ethical behavior in robots is challenging, as it’s difficult to encode complex human values into mathematical reward functions, leading to potential unintended consequences.
  • Summary: The ethical use of robots, inspired by concepts like Asimov’s Laws, is a significant concern. Reinforcement learning relies on reward functions designed by humans, which can be incomplete and fail to account for all possible scenarios, potentially leading to robots prioritizing programmed goals over safety or ethical considerations. This highlights the difficulty in translating human ethics into a format a robot can reliably follow.
Robot Failure and Self-Awareness
Copied to clipboard!
(00:11:38)
  • Key Takeaway: A major deficiency in current robots is their lack of self-awareness regarding the cause of their failures, unlike humans who can readily identify errors.
  • Summary: Unlike human children who can quickly understand why a task failed, robots often lack the ability to diagnose the source of their errors. While they can learn to avoid repeating actions that result in negative feedback, they don’t inherently understand if the failure was due to sensor calibration, environmental misestimation, or other factors. This lack of self-reflection hinders accelerated learning and robust performance in real-world scenarios.
Computer Animation to Robotics
Copied to clipboard!
(00:13:56)
  • Key Takeaway: The principles of physics-based animation in computer graphics are closely related to robotics, forming a foundation for generating realistic robot motion.
  • Summary: A background in computer animation, particularly physics-based animation, provides a strong foundation for robotics. This field involves using physics engines to generate realistic motion by calculating how joint torques, gravity, and contact forces affect a character’s poses. By replacing human animators with AI models that generate high-level commands, this process can be adapted to control virtual robots, which can then be deployed to the physical world.
Future of Home Robots
Copied to clipboard!
(00:15:42)
  • Key Takeaway: While fully capable home robots are still some years away, the open-sourcing of designs like Toddy makes building functional humanoid robots accessible now, with staged deployment expected.
  • Summary: Robots like Toddy, with its 3D-printed body and open-source designs, are available for individuals to build today. However, widespread adoption of highly capable home robots is projected to be a gradual process, similar to autonomous vehicles, with initial deployments likely focusing on specific tasks like laundry or dishwashing. The development of reliable, robust, and safe locomotion and manipulation systems is key to realizing this future.