Russ Tedrake: Underactuated Robotics, Control, Dynamics and Touch | Lex Fridman Podcast #114

Russ Tedrake joins Lex Fridman to explore the elegance of letting physics do the work in robotics, from passive dynamic walkers to barefoot running, which Tedrake practices on a near-marathon daily commute. He recounts MIT's DARPA Robotics Challenge journey, including the famous Atlas fall getting out of the car. The conversation digs into why contact and manipulation are so hard, the paradoxes of rigid-body physics, and how soft robotics and tactile sensing could change everything. They close on fleet learning, home robots for aging in place, fears of robots, and book recommendations.

• Tedrake calls passive dynamic walkers, powered only by gravity with no motors or batteries, the most beautiful robot motion he has ever seen.
  00:04:39
• He describes a dead fish tied behind a rock that begins swimming upstream purely from mechanics resonating with the vortex street, no brain or control.
  00:11:52
• Tedrake reveals he lost a finger in the early 2000s and his remaining bone remodeled itself to handle new contact.
  00:15:33
• Three days before the DARPA virtual competition the simulator suddenly got faster, forcing his team to write a new QP solver overnight to keep the robot from falling.
  00:52:25
• Atlas fell on day one of the physical competition because someone forgot to uncheck a box, leaving the driving controller fighting the balance controller while exiting the car.
  01:03:24
• A Black Mirror episode changed how a room of MIT faculty reacted to his Boston Dynamics videos, with a look of horror replacing prior excitement.
  01:18:58
• Tedrake argues robots are built wrong: they are afraid of touching the world and should be soft and contact-embracing rather than rigid.
  02:04:22

• Tedrake commutes ~12 miles to MIT and varies it by running barefoot, biking, rollerblading, or rowing a boat up the Charles River.
  00:27:57
• He retrained his running gait to about 180 steps per minute by running to songs at that beat.
  00:24:51
• In 1997-2004 the best walking robot was Honda's ASIMO, a marvel of engineering that walked cautiously with bent knees and was not energy efficient.
  00:07:14
• Tedrake warns that deep learning makes things easy to get working without rigorous thinking, and worries the next generation will lose that discipline.
  00:39:51
• The team faced a 'big robot little car problem': a 400-pound Atlas barely fit in the Polaris vehicle, forcing it to straddle the steering column.
  01:01:50
• He explains the rigid-body contact paradox: a four-legged table's friction forces are mathematically indeterminate, making contact hard to simulate.
  01:40:26
• Enveloping an egg with soft distributed contact lets a robot hold it without crushing it, unlike two hard point contacts.
  02:04:54
• Perception scientist Ted Adelson was struck by how much his kids learned by touching and licking everything, convincing him to study tactile sensing.
  02:25:09
• Tedrake describes a 'black swan generator' for rare-event simulation to find the rare but highly impactful ways a robot can fail.
  01:57:37
• Toyota Research Institute also uses AI and machine learning to discover new materials for car batteries.
  02:00:14