154. Can Robots Get a Grip?

Summary

Podcast Summary: People I (Mostly) Admire – Episode 154: Can Robots Get a Grip?

Introduction People I (Mostly) Admire episode 154, titled "Can Robots Get a Grip?", features an engaging conversation between host Steve Levitt and UC Berkeley robotics professor Ken Goldberg. Released on March 29, 2025, the episode delves deep into the intricacies of robotics, exploring the challenges, breakthroughs, and future prospects of robotic technology. Skipping over advertisements and non-content segments, this summary captures the core discussions, insights, and conclusions drawn during their conversation.

Ken Goldberg’s Journey into Robotics Ken Goldberg begins by sharing his early experiences with robotics, influenced by his father's attempt to build a robot for practical purposes in a chrome plating company.

Ken Goldberg [02:29]: "When I was a kid, I was really into rockets, models, building things like that. And my dad ran this chrome plating company… he wanted to build a machine robot that would do this dirty work."

Despite his father's creativity, the project never fully materialized due to practical challenges.

Ken Goldberg [03:23]: "I don’t think it ever really worked. My father was a great tinkerer, very creative, but he had a limited attention span, so he abandoned projects like that."

This early exposure set the foundation for Goldberg’s lifelong pursuit in robotics, particularly focusing on the problem of robotic grasping.

Defining a Robot: Beyond Humanoid Forms A significant portion of their discussion centers on what fundamentally defines a robot. Goldberg emphasizes that robots need not mimic human appearance but should be programmable machines capable of performing useful tasks in the physical world.

Ken Goldberg [04:22]: "A robot is a machine that's programmable, that moves in the physical world, but does something interesting and useful."

He critiques the prevailing obsession with humanoid robots, suggesting that functional design should take precedence over aesthetics.

Ken Goldberg [04:49]: "It's very compelling to have something that does have some form factor of a human humanoid. Yeah. And that is super popular right now."

The Challenges of Robotic Grasping Goldberg delves into the complexities of enabling robots to grasp objects, a task deceptively simple for humans but profoundly challenging for machines.

Ken Goldberg [07:02]: "Anytime you want to put something together, assemble it, you have to pick up the parts. It's very counterintuitive because it's much, much harder than people think."

He discusses his dissertation work on using parallel jaw grippers and the mathematical theories developed to enhance grasping efficiency without relying heavily on sensors.

Ken Goldberg [09:08]: "What I found was that there was this beautiful geometric way to essentially constrain the shape of any polygonal object so that it would come out in a unique final range."

Despite theoretical successes, practical implementations revealed vulnerabilities to minor errors and unpredictable factors like friction.

Ken Goldberg [11:01]: "The issue is there's very small errors and factors like friction that are hard to model. And when those get violated, the assumptions get violated, then things don't always work out as you hoped."

The Intersection of Robotics and Computer Vision The conversation transitions to the advancements in computer vision and its impact on robotics. Goldberg acknowledges the breakthroughs initiated by projects like ImageNet but points out that visual perception alone isn't sufficient for tasks requiring physical interaction.

Ken Goldberg [17:51]: "It's not completely nailed, I would say, but it was a breakthrough for sure."

He contrasts the perfect information environment of games like chess, where AI excels, with the unpredictable nature of the real world, highlighting the vast data requirements for robots to achieve similar proficiency.

Ken Goldberg [26:48]: "But if you compare that to the large language model, that's 1.2 billion hours. But that 10,000 hours is approximately a year. That means that we have so far accumulated one year. To get to the level of the large language models, that would take us 100,000 years."

DexNet: Bridging the Data Gap with Simulation To address the massive data gap, Goldberg introduces DexNet, a project that leverages simulation and domain randomization to enhance robotic grasping capabilities.

Ken Goldberg [29:55]: "We were able to generate a very large data set of three dimensional objects and grasps on objects and use that to train a neural network, but added noise so it was more realistic."

By incorporating random perturbations, DexNet trained robots to execute grasps that are robust to real-world imperfections, achieving impressive success rates.

Ken Goldberg [33:21]: "We were getting like well over 90% success rates."

This approach mirrors the "bitter lesson" in AI, where model-free, data-driven methods often outperform handcrafted models.

Ken Goldberg [24:00]: "This has been a bitter lesson for most researchers and academics… these models maybe don't work as well as this method that just sort of bubbles up out of magic."

Autonomous Vehicles: A Comparative Analysis Goldberg compares robotics to autonomous vehicles, explaining why companies like Waymo have succeeded over Tesla despite Tesla's vast data accumulation.

Ken Goldberg [37:14]: "Waymo is very successful. They have their cars running right, and they're actually very low accident rate."

He attributes Waymo's success to controlled data collection and a multifaceted sensory approach, contrasting it with Tesla's end-to-end learning philosophy, which struggles despite having more data.

Ken Goldberg [38:22]: "It's a different philosophy. I find this surprising because Elon Musk… is very good at good old fashioned engineering."

Future of Robotics: Complementarity Over Substitution When discussing the future interaction between humans and robots, Goldberg advocates for a complementary relationship where robots augment human capabilities rather than replace them.

Ken Goldberg [53:50]: "I’m 100% in on complementarity. This idea of augmenting our intelligence, our skills is so valuable."

He envisions a future where robots enhance tasks such as surgery, providing precision while humans oversee and control critical aspects.

Robotics in Art and Personal Endeavors Beyond his technical work, Goldberg shares his passion for art, illustrating how his creative pursuits influence his scientific endeavors and vice versa.

Ken Goldberg [45:18]: "And that's when we hit on, oh, have it garden. Because that was the last thing I would think people would want to do, because gardening is such a visceral."

His interdisciplinary approach led to projects like Telegarden, where an internet-controlled robot tended to a garden, engaging over 100,000 participants globally.

Ken Goldberg [47:11]: "The central sculpture in the exhibition is what we call the Tree of Knowledge… etched with all kinds of questions from the history of the evolution of knowledge on one side."

Addressing the Singularity and AI Concerns Goldberg addresses common fears surrounding AI and the concept of the singularity, expressing confidence that humans will maintain control over robotic advancements.

Ken Goldberg [54:28]: "I do not think that's going to happen. I think we're going to still be very much in control."

He dismisses the notion of robots spiraling out of control, emphasizing the importance of human oversight and the complementarity between human and machine intelligence.

Conclusion Episode 154 of People I (Mostly) Admire offers a comprehensive exploration of the current state and future of robotics through Ken Goldberg’s expert lens. From the foundational challenges of robotic grasping to the philosophical implications of AI advancements, Goldberg provides nuanced perspectives that bridge technical intricacies with broader societal impacts. His insights underscore the importance of balanced expectations, interdisciplinary collaboration, and the enduring value of human-robot complementarity.

Notable Quotes:

Ken Goldberg [04:22]: "A robot is a machine that's programmable, that moves in the physical world, but does something interesting and useful."
Ken Goldberg [24:00]: "This has been a bitter lesson for most researchers and academics… these models maybe don't work as well as this method that just sort of bubbles up out of magic."
Ken Goldberg [53:50]: "I’m 100% in on complementarity. This idea of augmenting our intelligence, our skills is so valuable."

Timestamp Highlights:

[02:29] Ken Goldberg discusses his early robotics project with his father.
[04:22] Defining what constitutes a robot.
[09:08] Goldberg’s dissertation on robotic grasping without sensing.
[24:00] Introduction of the "bitter lesson" in AI.
[29:55] Explanation of DexNet and domain randomization.
[37:14] Comparison between Waymo and Tesla’s autonomous vehicles.
[53:50] Advocacy for human-robot complementarity.
[54:28] Addressing fears about the AI singularity.

This comprehensive summary encapsulates the essence of the episode, providing listeners and non-listeners alike with valuable insights into the evolving landscape of robotics and AI.

Summary

Podcast Summary: People I (Mostly) Admire – Episode 154: Can Robots Get a Grip?

Ken Goldberg [02:29]: "When I was a kid, I was really into rockets, models, building things like that. And my dad ran this chrome plating company… he wanted to build a machine robot that would do this dirty work."

Despite his father's creativity, the project never fully materialized due to practical challenges.

Ken Goldberg [03:23]: "I don’t think it ever really worked. My father was a great tinkerer, very creative, but he had a limited attention span, so he abandoned projects like that."

This early exposure set the foundation for Goldberg’s lifelong pursuit in robotics, particularly focusing on the problem of robotic grasping.

Ken Goldberg [04:22]: "A robot is a machine that's programmable, that moves in the physical world, but does something interesting and useful."

He critiques the prevailing obsession with humanoid robots, suggesting that functional design should take precedence over aesthetics.

Ken Goldberg [04:49]: "It's very compelling to have something that does have some form factor of a human humanoid. Yeah. And that is super popular right now."

The Challenges of Robotic Grasping Goldberg delves into the complexities of enabling robots to grasp objects, a task deceptively simple for humans but profoundly challenging for machines.

Ken Goldberg [07:02]: "Anytime you want to put something together, assemble it, you have to pick up the parts. It's very counterintuitive because it's much, much harder than people think."

He discusses his dissertation work on using parallel jaw grippers and the mathematical theories developed to enhance grasping efficiency without relying heavily on sensors.

Ken Goldberg [09:08]: "What I found was that there was this beautiful geometric way to essentially constrain the shape of any polygonal object so that it would come out in a unique final range."

Despite theoretical successes, practical implementations revealed vulnerabilities to minor errors and unpredictable factors like friction.

Ken Goldberg [11:01]: "The issue is there's very small errors and factors like friction that are hard to model. And when those get violated, the assumptions get violated, then things don't always work out as you hoped."

Ken Goldberg [17:51]: "It's not completely nailed, I would say, but it was a breakthrough for sure."

Ken Goldberg [26:48]: "But if you compare that to the large language model, that's 1.2 billion hours. But that 10,000 hours is approximately a year. That means that we have so far accumulated one year. To get to the level of the large language models, that would take us 100,000 years."

Ken Goldberg [29:55]: "We were able to generate a very large data set of three dimensional objects and grasps on objects and use that to train a neural network, but added noise so it was more realistic."

By incorporating random perturbations, DexNet trained robots to execute grasps that are robust to real-world imperfections, achieving impressive success rates.

Ken Goldberg [33:21]: "We were getting like well over 90% success rates."

This approach mirrors the "bitter lesson" in AI, where model-free, data-driven methods often outperform handcrafted models.

Ken Goldberg [24:00]: "This has been a bitter lesson for most researchers and academics… these models maybe don't work as well as this method that just sort of bubbles up out of magic."

Ken Goldberg [37:14]: "Waymo is very successful. They have their cars running right, and they're actually very low accident rate."

He attributes Waymo's success to controlled data collection and a multifaceted sensory approach, contrasting it with Tesla's end-to-end learning philosophy, which struggles despite having more data.

Ken Goldberg [38:22]: "It's a different philosophy. I find this surprising because Elon Musk… is very good at good old fashioned engineering."

Ken Goldberg [53:50]: "I’m 100% in on complementarity. This idea of augmenting our intelligence, our skills is so valuable."

He envisions a future where robots enhance tasks such as surgery, providing precision while humans oversee and control critical aspects.

Robotics in Art and Personal Endeavors Beyond his technical work, Goldberg shares his passion for art, illustrating how his creative pursuits influence his scientific endeavors and vice versa.

Ken Goldberg [45:18]: "And that's when we hit on, oh, have it garden. Because that was the last thing I would think people would want to do, because gardening is such a visceral."

His interdisciplinary approach led to projects like Telegarden, where an internet-controlled robot tended to a garden, engaging over 100,000 participants globally.

Ken Goldberg [47:11]: "The central sculpture in the exhibition is what we call the Tree of Knowledge… etched with all kinds of questions from the history of the evolution of knowledge on one side."

Ken Goldberg [54:28]: "I do not think that's going to happen. I think we're going to still be very much in control."

He dismisses the notion of robots spiraling out of control, emphasizing the importance of human oversight and the complementarity between human and machine intelligence.

Notable Quotes:

Ken Goldberg [04:22]: "A robot is a machine that's programmable, that moves in the physical world, but does something interesting and useful."
Ken Goldberg [24:00]: "This has been a bitter lesson for most researchers and academics… these models maybe don't work as well as this method that just sort of bubbles up out of magic."
Ken Goldberg [53:50]: "I’m 100% in on complementarity. This idea of augmenting our intelligence, our skills is so valuable."

Timestamp Highlights:

[02:29] Ken Goldberg discusses his early robotics project with his father.
[04:22] Defining what constitutes a robot.
[09:08] Goldberg’s dissertation on robotic grasping without sensing.
[24:00] Introduction of the "bitter lesson" in AI.
[29:55] Explanation of DexNet and domain randomization.
[37:14] Comparison between Waymo and Tesla’s autonomous vehicles.
[53:50] Advocacy for human-robot complementarity.
[54:28] Addressing fears about the AI singularity.

This comprehensive summary encapsulates the essence of the episode, providing listeners and non-listeners alike with valuable insights into the evolving landscape of robotics and AI.

wavePod

Powered by Wave AI

Summary

Summary