Tedrake, Russ
▶ Watch on YouTubeCore ideas
Tedrake opens with passive dynamic walkers (Steve Collins and Andy Ruina's Cornell machines) as his favorite example of beautiful motion: legged mechanisms powered purely by gravity rolling down a ramp, with no motors, batteries, or controllers, achieving more natural-looking gait than powered humanoids like Honda's ASIMO. He explains the underlying principle via a "rimless wheel" model, where energy lost at each foot-strike and energy gained rolling downhill settle into a stable periodic motion, and frames this as a challenge to a control philosophy that tries to use motors to "erase the dynamics of the world" and impose simple, mathematically tractable behavior instead of letting physics do most of the work. He extends this idea with a striking biomechanics experiment: a dead rainbow trout, tethered behind a cylinder in a water tunnel to recreate the vortex street real fish exploit to conserve energy swimming upstream, will itself begin "swimming" upstream purely through passive resonance between its body mechanics and the vortex pattern — with no brain or muscles involved — illustrating how much apparent behavioral intelligence can reside in mechanics rather than control/computation. He generalizes this into a broader claim that mechanics and control aren't cleanly separable (his own bone remodeled around an old finger injury, showing "mechanical" adaptation independent of any nervous-system learning), and offers running barefoot (which he does daily on his ~12-mile commute) as a personal illustration: removing shoes provides immediate proprioceptive feedback that corrects bad gait mechanics (heel-striking, overstriding) before they cause long-term injury, versus shoes' delayed-feedback tendency to mask harmful movement patterns.
A major theme is the case for mathematical rigor against the pull of "it just works" deep learning: Tedrake argues that clear, precise control-theoretic thinking (contrasted with the fuzzier, statistics-driven intuitions common in machine learning) remains essential even in an era where deep learning "makes it easy to get things to work without thinking as critically," and that well-understood tools (like matrix inversion) shouldn't be replaced by learned approximations just because learning is available — while affirming that learning is essential where hand-designed models genuinely fail. He connects this to a critique of "state" itself: classical Lagrangian mechanics gives you a precise but often useless state representation (e.g., the "state" of a shirt being buttoned, or of 100 chopped onion pieces), and the real engineering challenge is finding a lower-complexity, task-relevant latent state representation that supports fast decision-making — a problem he sees as central to modern manipulation research, where perception (increasingly deep-learning-based) has to be integrated without demanding full-state estimation of enormously complex configurations.
The episode's most detailed technical thread is Tedrake's account of contact mechanics in robotics. He explains that rigid-body contact leads to genuine physical indeterminacy (e.g., a four-legged table's normal forces at each leg are underdetermined by basic conservation laws, which in turn makes friction forces, and thus which way a pushed table will slide, unpredictable from the model alone), producing mathematical "differential inclusions" rather than clean differential equations, and discontinuities whenever contacts form, break, or transition between sticking and slipping. This is why simulating and controlling contact-rich tasks (grasping, walking transitions between heel-strike and toe-off) is fundamentally harder than stabilizing systems like fighter jets, which can rely on locally linear, continuous dynamics — a distinction he argues is often mischaracterized as "contact is just hard to simulate" when the deeper issue is the rigid-body assumption's inherent paradoxes. He describes Drake (the physics/optimization/systems-modeling framework he built, now central to Toyota Research Institute's robotics work) as addressing this with rigorously validated multi-body physics and a systems-modeling philosophy (each component must formally declare its state/parameters, unlike ROS's more ad hoc "Mr. Potato Head" composability) that supports both simulation and formal controller verification. He advocates soft robotics and tactile sensing as a partial way out: soft, compliant contact surfaces turn brittle point-contacts into smoother, more forgiving patch contacts (illustrated by picking up an egg with distributed rather than point pressure), reduce discontinuity, and conveniently double as rich tactile sensors — solving the problem that hand/finger contact typically occludes head-mounted vision sensors at exactly the moment visual feedback matters most.
Tedrake recounts the DARPA Robotics Challenge as formative: his MIT team wrote a custom fast QP solver ("FastQP," built overnight with student Frank Permenter under real deadline pressure after simulator speed changed days before competition) using warm-starting to keep a real-time balance controller fast enough to prevent the 400-lb Atlas robot from literally falling over, and the team's most dramatic public failure occurred during a scripted car-egress task when an un-disabled driving controller conflicted with the balance controller in a rare, off-script contact configuration (torso briefly touching the seat) that the state estimator had never been designed to handle — a lesson he says shaped his thinking about testing, monte-carlo-style "rare event" / "black swan" simulation search for corner cases, and fleet learning (the idea that no single robot can visit enough of its own possible states, but a fleet collectively might) as future directions, alongside acknowledging that his team's slower testing cadence, versus the winning KAIST team's faster and more disciplined testing loop with two physical robots, was decisive. On public perception, he attributes fear of humanoid/legged robots more to cultural conditioning (citing the US's Terminator-saturated media diet versus Japan's Astro Boy) than any innate response, argues current robots don't warrant existential fear because the anthropomorphized capability gap is still large, and — echoing Rodney Brooks — frames the future less as "robots versus humans" and more as increasing human-robot integration (smartphones, pacemakers, eventual cognitive implants), locating the more pressing risks in the human-social effects of technology (e.g., how online interaction changes the value placed on in-person connection) rather than a robot/human battle line.
Excerpts
On the dead-fish experiment as evidence mechanics can substitute for control:
“They took a dead fish, they put a string that tied the mouth of the fish to the rock... and asked what would that dead fish do when it was hanging out behind the rock. What you'd expect is it sort of flopped around like a dead fish, until something sort of amazing happens: the dead fish basically starts swimming upstream — completely dead, no brain, no motors, no control — but somehow the mechanics of the fish resonate with the vortex street and it starts swimming upstream.”
On why rigid-body contact creates genuine physical indeterminacy:
“If I have a four-legged table and they were perfectly machined to be exactly the same height... the basic conservation laws don't tell me — there are many solutions for the forces that the ground could be putting on my legs that would still result in the table not moving... if you don't know what the normal force is on the four legs and you push the table, then you don't know what the friction forces are going to be.”
On why finding the right task-relevant state representation matters more than raw state complexity:
“If the first step of me trying to button my shirt is to estimate the full state of my shirt... that's just not the right specification. There are aspects of the state that are very important to the task, there are many that are unobservable and not important to the task — it begs new questions about state representation.”