Lex Fridman PodcastRuss Tedrake: Underactuated Robotics, Control, Dynamics and Touch | Lex Fridman Podcast #114
CHAPTERS
- 0:00 – 4:01
Lex sets the stage: Russ Tedrake’s work, barefoot running, and the scope of the conversation
Lex introduces Russ Tedrake’s background (MIT + Toyota Research Institute) and frames the discussion around control in underactuated, stochastic, hard-to-model robotic systems. He also previews themes that recur later: energy-efficient locomotion, rigorous thinking, and the connection between human movement and robotics.
- •Russ’s roles at MIT and Toyota Research Institute (TRI)
- •Focus on control in underactuated, stochastic, hard-to-model settings
- •Preview of topics: DARPA Robotics Challenge, locomotion, barefoot running, touch/manipulation
- •Lex’s podcast format and sponsor mentions (ads at the start)
- 4:01 – 9:41
Passive dynamic walking: letting physics do the work
Russ describes the beauty of passive dynamic walkers—robots that can walk down a ramp powered only by gravity, without batteries or controllers. The discussion contrasts “fighting” dynamics with motors versus embracing natural dynamics for efficiency and grace.
- •Passive dynamic walkers as an aesthetic and scientific milestone
- •Rimless wheel and compass-gait as simple models of walking
- •Stability of periodic gaits emerging from energy balance
- •ASIMO/complex robots: impressive but energy-inefficient when over-controlled
- •The ‘zen’ idea: cooperate with physics instead of canceling it
- 9:41 – 13:34
Animal movement and the dead-fish experiment: mechanics vs brain
Russ shares a favorite biomechanics story: trout “surfing” vortices behind rocks to save energy, culminating in the striking result that even a dead fish can swim upstream via passive resonance. The chapter highlights how evolution and mechanics shape efficient behavior beyond active control.
- •Rainbow trout gait changes behind rocks by exploiting vortex streets
- •Key insight: passive body dynamics can generate ‘smart’ behavior
- •Clincher experiment: dead fish exhibits upstream motion via resonance
- •Evolution optimizes more than efficiency (survival constraints first)
- •Animals adapt both neural control and body mechanics over time
- 13:34 – 33:02
Control vs mechanics, biped evolution, and why Russ runs barefoot
They unpack the blurry boundary between passive mechanics and active control, then pivot to human bipedalism and what bipeds are good for in robotics. Russ and Lex connect this to barefoot running as a feedback-rich way to learn efficient, injury-resistant gait mechanics.
- •Definition: mechanics (passive dynamics) vs control (active energy injection)
- •Why bipeds matter: world built for humans, but also a hard control challenge
- •Skepticism about simplistic evolutionary stories; randomness matters
- •Barefoot running as immediate feedback to avoid damaging gait patterns
- •Practical tips: start slowly; minimal shoes vs barefoot; proprioception and foot sensing
- 33:02 – 44:05
Rigorous thinking in the deep learning era: models, simplicity, and representation
Russ argues that deep learning can make it easy to get results without cultivating rigor, and that education should preserve precise reasoning. They discuss why simple models matter, how ‘complexity’ isn’t just state dimension, and why representation is central—especially in contact-rich robotics.
- •Control theory’s clarity vs ‘maybe/what-if’ conversations
- •Deep learning’s power—and its tendency to reduce incentives for rigor
- •Why simplifying to toy models can be a feature (Newton vs ‘weights of a net’)
- •Parsimonious explanations vs arbitrarily complex function approximators
- •Representation matters: contact creates combinatorial mode switches that demand new abstractions
- 44:05 – 50:49
DARPA Robotics Challenge: disaster response, Atlas, and the real meaning of robustness
Russ recounts MIT’s journey through the DARPA Robotics Challenge—from the motivation (Fukushima/disaster response) to competing for the right to use Atlas. He emphasizes the gap between making a system work once and making it robust enough to deploy under uncertainty and degraded communications.
- •DRC origins: disaster response, degraded comms, semi-autonomy goals
- •Two tracks: hardware teams vs software teams (virtual challenge)
- •Atlas as a heavy, expensive humanoid platform
- •Competition uncertainty: locomotion vs perception vs autonomy emphasis
- •Hard lesson: robustness requires a fundamentally different engineering/testing mindset
- 50:49 – 1:01:05
A solver written overnight: real-time optimization under pressure
Russ shares a high-stakes virtual-competition moment: when the simulator suddenly sped up, their controller had to match real-time performance. The team responded by creating a fast QP layer (warm-start-like behavior) to keep Atlas stable, illustrating how latency can directly cause catastrophic falls.
- •Cloud simulation architecture: controller machine vs physics machine
- •Real-time constraints driven by teleoperation and human interaction
- •Optimization-in-the-loop control: QP speed as a stability bottleneck
- •Emergency engineering: create a lightweight ‘Fast QP’ fallback system
- •Key idea: warm-start and exploiting similarity across consecutive time steps
- 1:01:05 – 1:07:15
Why robots fell at DRC: checklists, state estimation surprises, and the ‘big robot, little car’ problem
They dissect failure modes in dramatic DRC falls, highlighting that not all failures were “bad AI”—sometimes batteries or operational constraints were decisive. Russ tells MIT’s notorious fall while exiting a Polaris vehicle: a missed checklist step and an unmodeled configuration (butt-on-seat, feet-in-air) broke planning and state estimation.
- •Failures can be mundane: waiting drained batteries, leading to topples
- •Hardest task: egressing the Polaris vehicle in constrained geometry
- •Checklist discipline (NASA-style) and the cost of one missed step
- •Off-script states: planning + estimation can collapse in unanticipated configurations
- •Broader theme: robots are strong in scripted contact regimes, weak with incidental whole-body contact
- 1:07:15 – 1:18:39
From RoboCup to UFC robots: contact, game theory, and ‘rule exploitation’
Lex pushes the playful question of when robots will beat humans in physical sports or fighting. Russ argues that progress depends as much on competition ecosystems as on technical difficulty, and warns that robots might win by exploiting rules in ways that look nothing like human play.
- •Martial arts as ‘the art of contact’ and why it’s still hard for robots
- •Contact-rich decision-making requires fast reasoning and robust models
- •Competition momentum matters: RoboCup has decades of infrastructure
- •Physical contests raise ‘fairness’ issues due to robot strength/actuation
- •Robots may optimize to the rules, not to human-like behavior
- 1:18:39 – 1:34:00
Black Mirror and robot fear: culture, anthropomorphism, and real near-term risks
Russ explains how public perception shifted after Black Mirror’s robot-dog imagery, even among MIT faculty. He argues fear is partly cultural (Terminator vs Astro Boy), and reframes the risk: today’s biggest societal impacts come from integrated tech systems (phones, networks), not autonomous killer quadrupeds.
- •Black Mirror effect: changing emotional interpretation of Boston Dynamics videos
- •Why people anthropomorphize legged robots and over-attribute intelligence
- •Cultural narratives shape fear (US sci-fi vs Japan’s friendly robots)
- •Near-term harms can be software-based (misuse, surveillance, drones)
- •Long-term view: humans and robots will likely co-evolve into ‘robot people’
- 1:34:00 – 1:46:58
What is a robot, and what makes control hard: contact and discontinuities
Russ offers a pragmatic definition of robots—computation plus mechanical work—while noting that mature successes stop being called “robots.” He then explains why contact makes control fundamentally harder than smooth dynamical systems: discontinuities, mode switches (stick/slip), and combinatorial contact configurations.
- •Robots ‘disappear’ as they succeed (dishwashers, factory arms)
- •Control differs dramatically across robot types; Russ focuses on dynamic interaction
- •Contact introduces non-smooth dynamics and discontinuous transitions
- •Rigid-body + friction modeling creates indeterminacies (e.g., four-legged table forces)
- •Contact configurations explode combinatorially (walking feet, dexterous hands, object interactions)
- 1:46:58 – 1:54:53
Simulating robots with Drake: optimization, system modeling, and contact physics
Russ describes Drake as more than a simulator: it’s an optimization library, a system modeling framework, and a multibody physics engine with extensive testing. The aim is not just to demo a task once (like dish loading), but to build the engineering foundation for verification, validation, and deployable reliability.
- •Drake’s three pillars: optimization layer, system modeling language, multibody physics engine
- •Contrast with ROS: more explicit system declarations enable design/verification workflows
- •Heavy engineering investment: correctness tests (spinning tops, satellites, regression suites)
- •Simulation as a tool for product-level confidence, not just research demos
- •Key pain point: reproducing one-off hardware failures inside simulation for debugging and prevention
- 1:54:53 – 2:00:10
Finding corner cases: falsification, rare events, and fleet learning
They explore how to systematically discover “black swan” failures using accelerated testing, falsification algorithms, and rare-event simulation—ideas that become essential when real-world testing is slow and non-repeatable. Russ connects this to fleet learning: many robots collectively exploring state space the way billions of humans do socially.
- •Active experiment design (modern ‘curiosity’) to reduce uncertainty efficiently
- •Falsification/rare-event simulation to concentrate tests on failure regions
- •‘Black swan generator’: seek rare and high-impact failures, not just rare events
- •Monte Carlo testing limits: too much space, too little time, too many corner cases
- •Fleet learning: distributed experience across many robots changes what’s feasible to learn/validate
- 2:00:10 – 2:03:40
Home robotics: manipulation as the next breakout after drones and self-driving
Russ outlines TRI’s home-robotics focus and why manipulation is poised to ‘pop’ as a major field. They discuss product realities (safety vs usefulness), and Toyota’s motivation around helping people age in place—an emotionally and technically delicate application where reliability and human trust matter.
- •TRI portfolio: home robotics plus other AI-driven efforts (e.g., materials discovery)
- •Why home manipulation is hard yet promising (beyond ‘pick and place’)
- •Aging-in-place as a guiding application: assistance without disempowerment
- •Human tolerance for ‘learning systems’ when progress is visible (fleet improvement analogy)
- •Safety-critical contrast: cars vs household tasks, and why the home may allow faster iteration
- 2:03:40 – 2:07:25
Soft robotics and tactile sensing: making contact smoother and more informative
Russ argues robots are ‘built wrong’ when only fingertips/feet are compliant but the rest is rigid metal. Softness can reduce discontinuities in contact, distribute forces (e.g., picking up an egg), and enable embedding tactile sensors—crucial when the robot’s own hand occludes vision at the moment contact matters most.
- •Soft contact turns point contacts into patch contacts (better stability and force distribution)
- •Softness makes outcomes smoother and less catastrophically sensitive to small errors
- •Embedding sensors in compliant skins enables rich tactile perception
- •Vision occlusion problem: when hands move in, head cameras lose visibility
- •Soft robotics + tactile sensing as a path toward safe, capable human-proximate manipulation
- 2:07:25 – 2:20:36
Underactuated robotics: why you can’t control everything (and why that’s the point)
Russ defines underactuation as having fewer actuators than degrees of freedom—and argues the world forces underactuation on us (humans included). He explains why optimal control/optimization has been the most productive tool for managing underactuated dynamics in legged robots and beyond.
- •Definition: degrees of freedom vs actuators; underactuated means a mismatch
- •Humans are underactuated (e.g., center-of-mass motion in flight)
- •Walking can look ‘almost fully actuated’ in contact, until constraints matter
- •Manipulation is inherently underactuated relative to object goals (controlling the cup, not just joints)
- •Optimization/numerical optimal control as a practical engine for underactuated control design
- 2:20:36 – 2:28:52
Touch, embodiment, and learning: why robots need active physical exploration
Lex and Russ zoom out from mechanics to meaning: touch is essential not only for safety but for learning and human connection. Russ cites how infants learn by touching/licking everything, reframing robotics as an active experimentation problem rather than passive perception alone.
- •Touch enables safe, meaningful physical collaboration with people
- •Design matters: soft, ‘Baymax-like’ robots may invite trust and teaching
- •Anthropomorphism as a lever for human-robot interaction quality
- •Active perception: learning by intervening is fundamentally different than observing
- •Robots need tactile feedback to act reliably in the occluded, contact-critical phase of manipulation
- 2:28:52 – 2:40:00
Books, reading deeply, and teaching as a forcing function for clarity
Russ recommends books spanning AI geopolitics and big-picture history, then emphasizes a meta-skill: learning to read deeply and selectively amid information overload. He also describes teaching as a discipline that forces stronger internal models through explanation and questioning.
- •Recommendations: AI Superpowers (Kai-Fu Lee), Sapiens/Homo Deus (Harari), The Black Swan (Taleb)
- •A ‘controversial’ favorite: How to Read a Book (Mortimer Adler)
- •Core idea: a great book is a cross-time dialogue with the author
- •Information overload: prioritize a few foundational works and study them deeply
- •Teaching as a method to refine understanding via repeated explanation and audience questions
- 2:40:00 – 2:47:09
Advice to young people and the meaning in doing hard things
Russ advises young people to cultivate deep, critical thinking and build intuition by doing—math, engineering, and taking things apart. The conversation closes with reflections on being challenged (especially early at MIT), loving the learning journey, and finding meaning in continuously discovering what you don’t yet understand.
- •Winners in a noisy world: those who can think deeply and critically
- •Start math early; build things; learn by disassembly and reconstruction
- •Expect a long journey—no one masters everything
- •Personal story: arriving at MIT, feeling overwhelmed, but energized by the challenge
- •Meaning emerges from the ongoing process of learning and connecting puzzle pieces
- 2:47:09 – 2:48:46
Closing remarks and sponsor outro
Lex closes the conversation, thanks Russ, and returns briefly to sponsor acknowledgments. He ends with a Neil deGrasse Tyson quote about why human exploration inspires the public more than purely robotic discovery.
- •Lex thanks Russ and reflects on robotics as ‘early days’ of a key future field
- •Sponsor reminders and ways to support the podcast
- •Neil deGrasse Tyson quote: public inspiration and the human element in exploration
- •Outro framing: robots matter, but narratives and human participation drive engagement
