For the next few weeks, the Actuator newsletter will feature Q&As with leading experts in robotics. Don't miss out—subscribe here for updates.
Part 1: Matthew Johnson-Roberson, Carnegie Mellon University
Part 2: Max Bajracharya and Russ Tedrake, Toyota Research Institute
Part 3: Dhruv Batra, Meta
Part 4: Aaron Saunders, Boston Dynamics
Ken Goldberg serves as a professor and the William S. Floyd Jr. Distinguished Chair in Engineering at UC Berkeley. He is also the co-founder and chief scientist at the robotics parcel sorting startup Ambidextrous and an IEEE fellow.
What Roles Will Generative AI Play in the Future of Robotics?
2023 is set to be a landmark year as generative AI begins to reshape the field of robotics. Large language models, such as ChatGPT, enable seamless communication between robots and humans, allowing for natural language interactions. Words, from "chair" to "chocolate," have evolved over time to convey valuable concepts. Moreover, roboticists are leveraging large Vision-Language-Action models to enhance robot perception and control the movements of robotic arms and legs. With the need for extensive data, labs worldwide are collaborating to share resources, resulting in promising advancements. While questions about generalization remain, the potential impact is significant.
An emerging focus area is "Multi-Modal models," which can be understood in two contexts:
1. Input Modes: Combining different sensory inputs such as Vision and Language, with expansions now including Tactile and Depth sensing, as well as Robot Actions.
2. Actions from Input States: The ability to perform various actions in response to identical input conditions. This is common in robotics, as evidenced by the many methods available for grasping objects. Standard deep learning models often average these grasping techniques, which can lead to poor results. A breakthrough method called Diffusion Policies, developed by Shuran Song at Stanford, offers a promising way to maintain the variety of multi-modal actions.
What Are Your Thoughts on the Humanoid Form Factor?
Historically, I have been skeptical of humanoid and legged robots due to their potential for inefficiency. However, that perspective is shifting as I observe the latest advancements from companies like Boston Dynamics, Agility, and Unitree. Tesla's engineering capabilities position them well to develop scalable low-cost motors and gearing systems. Legged robots hold significant advantages over wheeled models as they can navigate obstacles such as steps, debris, and rugs more effectively. While bimanual (two-armed) robots are vital for numerous applications, I maintain that simple grippers will remain more reliable and cost-efficient than complex five-fingered robotic hands.
Following Manufacturing and Warehousing, What's Next for Robotics?
In light of recent wage settlements in unions, I anticipate a surge in robotics applications within manufacturing and warehouses. Progress in self-driving taxis, particularly in challenging environments like San Francisco, has been noteworthy. Yet, I remain skeptical about their long-term cost-effectiveness. In the realm of healthcare, researchers are investigating "Augmented Dexterity," a concept where robots enhance surgical capabilities by executing lower-level tasks, such as suturing.
How Far Are We from True General-Purpose Robots?
I do not foresee the arrival of true AGI or general-purpose robots in the immediate future. Among my peers, there is little concern regarding robots taking jobs or assuming control.
Will Home Robots Beyond Vacuums Become Popular in the Next Decade?
I predict that within the next decade, we will see cost-effective home robots capable of decluttering spaces by picking up clothes, toys, and trash, and sorting them into appropriate bins. Similar to today's vacuum cleaners, these robots may occasionally make mistakes, but the advantages for busy families and elderly individuals will outweigh potential downsides.
What Important Robotics Trend Deserves More Attention?
Robot motion planning remains a critical yet often overlooked topic in robotics. It involves controlling motor joints to facilitate tool movement and avoid obstacles. Many believe this issue has been resolved, but that is far from the truth.
Robot singularities pose a fundamental challenge across all robotic arms. Unlike futurist Ray Kurzweil's notion of an AI singularity, robot singularities refer to spatial points where a robot halts unexpectedly and requires manual reset from an operator. These singularities arise from the mathematical computations needed to translate desired motions of a gripper into the corresponding movements of each of the six robot joint motors. At specific positions, this conversion can become unstable, similar to a divide-by-zero error, prompting a mandatory reset.
For repetitive motions, meticulous fine-tuning can help avoid singularities. However, in the expanding scope of non-repetitive applications—like palletizing, bin-picking, order fulfillment, and package sorting—singularities frequently disrupt operations unpredictably (sometimes several times an hour). To tackle this challenge, I co-founded Jacobi Robotics, a startup implementing efficient algorithms guaranteed to circumvent singularities, significantly enhancing the reliability and productivity of robotic systems.
Ranking
