Why don't we have better robots yet? | Ken Goldberg

Episode Summary

Executive Summary: Ken Goldberg explains why robots still struggle with everyday manipulation: human-easy tasks like grasping, untangling, folding, and bagging are hard because of uncertainty in control, sensing, and physics. He shows how AI and simulation have improved warehouse picking and package sorting, while home robotics remains an open challenge.

Main Topics: Moravec’s Paradox and the home-robot gap (Priority: 5/5): Goldberg frames the central puzzle of robotics: tasks that seem simple to humans are often the hardest for robots, explaining why household robots remain limited despite decades of progress. Hardware simplicity vs. software complexity (Priority: 5/5): He argues that robot hands should be simple and reliable, but the real difficulty lies in software—especially dealing with uncertainty in control, perception, and physical interaction. Sensors and the limits of perception (Priority: 4/5): The talk compares cameras, LiDAR, and tactile sensors, noting that each improves robot understanding but still fails in important edge cases like shiny, transparent, or hidden surfaces. Warehouse automation and e-commerce (Priority: 5/5): Goldberg highlights e-commerce as a practical success area where robots can sort packages at scale, driven by demand for faster fulfillment and high human turnover in warehouses. Simulation, deep learning, and self-training robots (Priority: 5/5): He describes DexNet and related methods where robots learn grasping in simulation, enabling reliable picking of unseen objects and powering commercial systems. Toward more capable home manipulation (Priority: 4/5): Current research targets deformable-object tasks such as cable untangling, laundry folding, and bagging—showing progress but also how far robots still are from general household usefulness.

Key Arguments: Robots are still clumsy because manipulation requires solving uncertainty in control, sensing, and physics simultaneously. Simple grippers are often better than complex hands because reliability and cost matter more than human-like anatomy. LiDAR and tactile sensing help, but no sensor fully eliminates ambiguity in real-world scenes. Warehouse picking is one of the few areas where robotics has found strong product-market fit because the task is repetitive, high-volume, and economically valuable. Deep learning plus simulation can let robots train themselves to grasp objects they have never seen before. Home robotics remains difficult because everyday objects are deformable, cluttered, and variable, making tasks like folding and bagging much harder than they appear. Progress is real, but robots still need humans for many tasks robots cannot yet do.

Data Points: Years of research at UC Berkeley: 30 years - Goldberg says he has studied robots with his students for three decades. Warehouse machines in operation: 80 - Ambi Robotics machines deployed across the United States. Packages sorted per week: over 1 million - Scale of package sorting performed by Ambi Robotics systems. Cable untangling success rate: about 80% - Robot performance on tangled cable manipulation. Laundry folding speed in prior research: 3 to 6 folds per hour - Earlier robotic laundry-folding work Goldberg cites as too slow. Bagging success rate: about half the time - Robot performance on opening/manipulating bags. Talk length: 10 minutes - Goldberg says he will explain the fiction-reality gap in the next 10 minutes. Pandemic e-commerce trend: huge jump - Online ordering increased sharply during the pandemic, intensifying warehouse automation demand.

Pivotal Quotes: "What's easy for robots, like being able to pick up a large object, large heavy object, is hard for humans. But what's easy for humans, like being able to pick up some blocks and stack them, well, it turns out that is very hard for robots." — Ken Goldberg: Defines Moravec’s paradox and the core challenge of robotics. "We have incredible capabilities. We're very good at manipulation. But robots still are not." — Ken Goldberg: Summarizes the gap between human dexterity and robotic manipulation. "The robot would do this in simulation. It was almost as if the robot were dreaming about how to grasp things and learning how to grasp them reliably." — Ken Goldberg: Describes how DexNet trains grasping through simulated practice.

Implications: Robotics is advancing fastest where tasks are structured and economically urgent, especially logistics. But truly useful home robots will require major gains in perception, dexterity, and learning for messy real-world manipulation.