Vijay Kumar: One of the world's top roboticists, professor and Dean of Engineering at the University of Pennsylvania, former director of the GRASP lab, known for multi-robot systems, swarms, and micro aerial vehicles.
What this episode covers
Vijay Kumar traces his journey from building a 7,000-pound hexapod in graduate school to pioneering small agile quadrotors that cooperate in 3D formations. He explains how advances in computing and cheap IMUs (driven by car airbags) around 2007-2009 made onboard sensing and agile flight possible. The conversation covers true autonomy without GPS, communications, or human pilots, the role of learning versus model-based control, and why perception has advanced faster than learned action. Kumar argues autonomous flight is in some ways easier than driving, but flying cars face a fundamental battery energy and power density barrier. He closes on human-robot collaboration, weaponization risks of swarms, and advice for young engineers to embrace math, breadth, and the liberal arts.
Also referenced (named, not recommended)
Raspberry Pi
Raspberry Pi Foundation
“a high school kid today can pick up a Raspberry Pi kit and build us all the low levels functionality is all automated”— Vijay Kumar
iPhone
Apple
“I think 2007 was a tipping point the year of the iPhone the year of the cloud lots of things happen in 2007”— Vijay Kumar
The Jetsons
Hanna-Barbera
“they would look like something like we dream of when we think about flying cars yeah like the Jetsons The Jetsons yeah”— Vijay Kumar
Netflix
Netflix
“a lot of people have sci-fi inspired fears of walking robots like those from Boston Dynamics if you just look at shows on Netflix”— Lex Fridman
Big reveals from this episode
- Cheap IMUs for drones were enabled by airbag regulations requiring an accelerometer in every car, which drove down the price-to-performance ratio.
- True autonomy means navigating with no pilots, no communications, no base station, no GPS position, and no a priori map of the environment.
- On the action side of autonomy Kumar knows of no fielded systems that actually learn correct behavior end-to-end; only perception learning has truly succeeded.
- Going from 90% to 99% to 99.9% perception accuracy requires data growing along what Kumar believes is an exponential curve.
- In 2014 two percent of US electricity consumption went to data farms, making AI an energy-processing problem, not just an information one.
- Autonomous flight can use a guaranteed-safe trajectory (up, across, down) that has no easy equivalent for ground-bound self-driving cars.
- Lifting a small UAV vertically costs roughly 200 watts per kilo, while the entire human brain runs on under 80 watts.
Worth remembering
- Kumar's first robot was a roughly 7,000-pound hexapod with 18 hydraulic motors, each controlled by its own computer plus a 19th coordinating computer.
- The clock speeds on those 1980s control computers were about half a megahertz.
- The team demonstrated small UAVs forming and deforming 3D patterns in midair back in 2011, which was a major accomplishment then and standard now.
- An ant with one of its six legs removed keeps moving fine and may not even realize it lost a leg, illustrating individual robustness.
- The four-motor quadrotor configuration has at least a hundred-year history and was not invented by Kumar's lab.
- Rotor blades flap and bend rather than staying rigid, and ground effect plus ceiling suction and wall effects create microclimates that disturb UAVs.
- Kumar works with a nonprofit, Weave Robotics, delivering drugs and retrieving test samples in the Peruvian Amazon where rivers are the only highways.
- We can achieve level-five autonomy in a parking lot but not on the streets of Naples or Mumbai.
- Kumar may be the only TED speaker to show a math equation, after the curator told him he couldn't show math.