In the coming weeks, TechCrunch’s robotics newsletter, Actuator, will feature Q&As with some of the top people in robotics. Subscribe here for future updates.
Part 1: Matthew Johnson Roberson of CMU
part 2: Max Bajracharya and Russ Tedrake of Toyota Research Institute
Part 3: Dhruv Batra in Mehta
Part 4: Aaron Sanders of Boston Dynamics
Ken Goldberg is a professor at the University of California, Berkeley, the William S. Floyd Jr. Distinguished Chair in Engineering, and co-founder and principal scientist of a robotic package sorting startup. He is ambidextrous and a Fellow of the IEEE.
What role will generative AI play in the future of robotics?
Although the rumors started a little early, 2023 will be remembered as the year generative AI transformed robotics. Large-scale language models like ChatGPT allow robots and humans to communicate in natural language. Words have evolved over time to express useful concepts, from “chair” to “chocolate” to “charisma.” Robotics engineers have also discovered a great visual language.action The model facilitates robot recognition and can be trained to control the robot’s arm and leg movements. Training requires vast amounts of data, so labs around the world are now collaborating and sharing data. The results continue to emerge, and although there are still open questions about generalizability, the impact will be profound.
Another interesting topic is “multimodal models” in two senses of multimodal.
- Multimodal, combining different input modes such as visual and verbal. This has now been expanded to include tactile and depth sensing, as well as robotic actions.
- Multimodal in that it allows different actions depending on the same input state. This is surprisingly common in robotics. For example, there are many ways to grasp objects. Standard deep models “average” these grasp actions, which can produce very poor grasps. One of his very attractive methods for preserving multimodal actions is Diffusion Policies, currently developed by Shuran Song at Stanford University.
What do you think about the humanoid form factor?
I’ve always been skeptical of humanoids and legged robots because they can be too sensational and inefficient, but the latest humanoid and quadrupedal robots from Boston Dynamics, Agility, and Unitree have made me reconsider. Ta. Tesla has the engineering skills to develop low-cost motor and gear systems at scale. Legged robots have many advantages over wheels when navigating steps, debris, and rugs in homes and factories. Although two-handed (two-armed) robots are essential for many tasks, I still believe that a simple gripper is more reliable and cost-effective than his five-fingered robot hand.
What will be the next major category of robots after manufacturing and warehousing?
After the recent union wage settlement, I think we’ll see even more robots in manufacturing and warehouses than we do now. Recent advances in self-driving taxis have been impressive, especially in San Francisco, where driving conditions are more complex than in Phoenix. But I’m not convinced they are cost effective. For robot-assisted surgery, researchers are exploring “enhanced dexterity,” or the ability of robots to improve surgical skills by performing low-level subtasks such as suturing.
How far have true general-purpose robots evolved?
I don’t think we’ll see true AGI or general-purpose robots in the near future. No roboticist I know is worried about robots taking their jobs or taking over.
Will household robots (beyond vacuum cleaners) become commonplace within the next 10 years?
I predict that within the next 10 years, we’ll see affordable household robots that can pick up clothes, toys, trash, etc. from the floor and put them in the appropriate bins. Like today’s vacuum cleaners, these robots will occasionally make mistakes, but the benefits to parents and seniors will likely outweigh the risks.
What are some important robotics stories/trends that aren’t getting enough coverage?
Robot motion planning. This is one of the oldest subjects in robotics and is how to control the joints of motors to move robotic tools and avoid obstacles. Many people think that this problem has been resolved, but it is not resolved yet.
Robotic “singularity” is a fundamental problem for all robotic arms. They are very different from Kurzweil’s hypothetical point at which AI will surpass humans. A robot singularity is a point in space where the robot stops unexpectedly and must be manually reset by a human operator. The singularity arises from the calculations required to convert the desired linear motion of the gripper into the corresponding motion of each of his six robot joints and her motors. At certain points in space, this transformation becomes unstable (similar to a divide-by-zero error) and the robot must be reset.
For repetitive robot motions, singularities can be avoided by tedious manual fine-tuning of the robot’s repetitive motions to avoid encountering singularities. Once such a behavior is determined, it is repeated over and over again. However, singularities are common in a growing generation of applications where robot movements are non-repetitive, such as palletizing, bin picking, order processing, and package sorting. These are well-known fundamental problems because they disrupt the robot’s operation at unexpected times, often several times an hour. I co-founded a new startup, Jacobi Robotics. Implement efficient algorithms that are “guaranteed” to avoid singularities. This significantly increases the reliability and productivity of all robots.