Latest / Mind Cast / Faster Than Bolt | The 8.64-Second Sprint That Changed Physical AI

Transcript

Faster Than Bolt | The 8.64-Second Sprint That Changed Physical AI · Mind Cast

• 0:00
  SPEAKER_00: 8.64 seconds.
• 0:02
  SPEAKER_00: That's all it took.
• 0:03
  SPEAKER_00: 8.64 seconds for a humanoid robot, a machine built of aluminum, carbon fiber, and silicon to run 100 meters.
• 0:14
  SPEAKER_00: Usain Bolt, the greatest sprinter who ever lived, set his world record at 9.58 seconds back in 2009.
• 0:23
  SPEAKER_00: That record has stood for over 15 years.
• 0:26
  SPEAKER_00: No human being on Earth has come close to breaking it.
• 0:30
  SPEAKER_00: A robot just did, on a track in Beijing, in front of thousands of people.
• 0:36
  SPEAKER_00: A humanoid robot, two legs, two arms, upright posture, crossed the finish line faster than the fastest human who has ever lived.
• 0:47
  SPEAKER_00: Let that land for a second.
• 0:51
  SPEAKER_00: Welcome to Minecast.
• 0:53
  SPEAKER_00: I'm your host, Will, and today we are going somewhere I genuinely did not expect to go.
• 0:59
  SPEAKER_00: We are talking about the 2026 World Humanoid Robot Games, held at Beijing's National Speed Skating Oval.
• 1:08
  SPEAKER_00: 2,056 robots, 666 teams from 16 countries, 51 events, and a set of implications that reach far beyond athletics.
• 1:20
  SPEAKER_00: Here's my promise: by the time we're done, you won't just know what happened in Beijing, you'll understand why it matters, for the future of work, for geopolitics, for the global economy, and for what it means to be human in an age when machines outrun us.
• 1:38
  SPEAKER_00: The sprint record is not the most important story from Beijing.
• 1:42
  SPEAKER_00: The most important story is what happened after the robots crossed the finish line, and we're gonna get there.
• 1:49
  SPEAKER_00: Let's start with a revolution in how these machines are built.
• 1:53
  SPEAKER_00: The Architecture Revolution.
• 1:56
  SPEAKER_00: Cast your mind back to 2013, DARPA launched the Robotics Challenge.
• 2:01
  SPEAKER_00: The goal?
• 2:02
  SPEAKER_00: Build a humanoid robot that could navigate a disaster zone, open a door, turn a valve, drive a car.
• 2:10
  SPEAKER_00: The robots that competed were tethered to power cables, moving at speeds between 0.075 and 0.887 meters per second, slower than a brisk walk.
• 2:23
  SPEAKER_00: They fell over constantly, taking minutes to complete tasks a healthy adult would do in seconds.
• 2:30
  SPEAKER_00: Fast forward 12 years to the ice ribbon in Beijing, and we're looking at machines hitting a peak sprint velocity of 14.5 meters per second, roughly 32 miles an hour.
• 2:44
  SPEAKER_00: So what changed in those 12 years?
• 2:47
  SPEAKER_00: The answer is the shift from model-based control to deep reinforcement learning, or DRL.
• 2:54
  SPEAKER_00: Let me explain what that means because it's crucial.
• 2:57
  SPEAKER_00: The old approach, what engineers call ZMP or zero moment point control, gave a robot a pre-written rulebook for every possible movement.
• 3:07
  SPEAKER_00: Engineers modeled every joint, every force, every scenario, and wrote equations to handle it.
• 3:14
  SPEAKER_00: The robot didn't learn, it executed.
• 3:17
  SPEAKER_00: And if reality deviated from the model, a wet floor or a wind gust, the system could fail catastrophically.
• 3:25
  SPEAKER_00: Deep reinforcement learning is categorically different.
• 3:29
  SPEAKER_00: Instead of writing the rules, you create a simulation environment and let the robot learn.
• 3:35
  SPEAKER_00: You reward it for staying upright, penalize it for falling, and run millions of simulated steps over days or weeks.
• 3:44
  SPEAKER_00: Gradually, through trial and error at a scale no human engineer could manually program, the robot develops an intuitive sense of balance and coordination.
• 3:54
  SPEAKER_00: The old approach was like teaching someone to ride a bicycle by giving them a physics textbook.
• 4:00
  SPEAKER_00: The new approach is letting them fall 10 million times in simulation until their body just knows what to do.
• 4:08
  SPEAKER_00: And this shift didn't just improve the robots, it transformed them.
• 4:13
  SPEAKER_00: This brings us to the hardware side.
• 4:15
  SPEAKER_00: One of the key metrics researchers use is something called cost of transport, COT.
• 4:21
  SPEAKER_00: It's how efficiently a system moves a unit of mass over a unit of distance, miles per gallon, but for robots.
• 4:29
  SPEAKER_00: In the DARPA era, humanoid robots had a COT of between 1.5 and 3.0, three to six times worse than a healthy walking human.
• 4:40
  SPEAKER_00: By 2026, the leading Beijing platforms had brought that down to between 0.2 and 0.5, approaching human-level efficiency in just over a decade.
• 4:52
  SPEAKER_00: That improvement comes largely from a change in the actuators, the motors driving the joints.
• 4:58
  SPEAKER_00: Old systems used harmonic drives, precise but energy hungry, with high friction losses and limited dynamic responsiveness.
• 5:07
  SPEAKER_00: The new systems use quasi-direct drive actuators, QDD.
• 5:11
  SPEAKER_00: Force transmits directly, feedback is sensed immediately, and energy is recaptured during deceleration rather than wasted as heat.
• 5:21
  SPEAKER_00: Better software learning plus better hardware, that's what produced the machines we saw in Beijing.
• 5:27
  SPEAKER_00: The crash that revealed everything.
• 5:30
  SPEAKER_00: Now, here's where the story gets really interesting and counterintuitive.
• 5:35
  SPEAKER_00: The robots that won the sprint, the machines that crossed the finish line in 8.64 seconds faster than Usain Bolt, crashed.
• 5:44
  SPEAKER_00: They slammed into the barriers at the end of the track.
• 5:47
  SPEAKER_00: Multiple robots experienced broken joints, electrical shorting, and thermal events inside the joint housings.
• 5:54
  SPEAKER_00: Why?
• 5:55
  SPEAKER_00: Because they physically could not stop in time.
• 5:59
  SPEAKER_00: Stopping distance equals mass times velocity squared divided by twice the braking force.
• 6:05
  SPEAKER_00: A 70 kilogram robot moving at 14.5 meters per second requires enormous braking force to stop within a reasonable distance.
• 6:14
  SPEAKER_00: The braking systems on these robots simply weren't capable of handling that demand at those speeds, so they ran straight into the barriers.
• 6:23
  SPEAKER_00: Now you might be thinking, okay, that's an engineering problem.
• 6:27
  SPEAKER_00: They'll fix it.
• 6:28
  SPEAKER_00: And yes, they will.
• 6:30
  SPEAKER_00: But the researchers at Beijing were pointing to something deeper.
• 6:34
  SPEAKER_00: Dennis Hong from UCLA and Jonathan Hurst from Oregon State both made the same observation.
• 6:41
  SPEAKER_00: Speed on a flat track tells you almost nothing about real-world capability.
• 6:46
  SPEAKER_00: A 100-meter sprint tests one thing: maximum force in one direction on a flat, predictable surface.
• 6:54
  SPEAKER_00: No obstacles, no decisions, no adaptation.
• 6:58
  SPEAKER_00: Real environments have stairs, uneven floors, cluttered workspaces, unpredictable humans, objects of varying weight and fragility.
• 7:07
  SPEAKER_00: This is why the structure of the 2026 Games was so cleverly designed, because it wasn't just a sports competition, it was a two-track evaluation.
• 7:17
  SPEAKER_00: One track, 30 athletic events, including a standing high jump where the robot record hit 2.88 meters, eclipsing the human world record of 2.45 meters, set by Javier Sotomayor.
• 7:32
  SPEAKER_00: The other track, 21 scenario-based industrial tasks, automotive assembly, logistics, hotel service, disaster response.
• 7:42
  SPEAKER_00: And here's what the industrial tasks revealed.
• 7:45
  SPEAKER_00: The robots that dominated on the track were not necessarily the same robots that excelled in the scenario tasks.
• 7:53
  SPEAKER_00: Different design philosophies, different optimization targets, different trade-offs.
• 7:59
  SPEAKER_00: Consider the UBTech Walker S2, 52 degrees of freedom across its body, 22 in its hands alone, priced between$140,000 and$199,000 with a three-minute hot swap battery already deployed at BYD, Foxconn, and Neo.
• 8:20
  SPEAKER_00: That's not a sports robot, that's a labor robot, and it performs very differently on a 100-meter sprint than it does threading a component through an assembly jig.
• 8:31
  SPEAKER_00: The conventional wisdom in tech media has been faster equals better.
• 8:36
  SPEAKER_00: But what Beijing demonstrated with robots literally smashing into barriers as the evidence is that the benchmark we should actually be watching is not the sprint time, it's the industrial task score.
• 8:49
  SPEAKER_00: The sprint is the spectacle, the scenario tasks are the signal.
• 8:54
  SPEAKER_00: The geopolitical chess board.
• 8:56
  SPEAKER_00: What happened in Beijing isn't just a robotics story, it's a geopolitical story, essential context for anyone in manufacturing, technology, or policy.
• 9:07
  SPEAKER_00: Here's the headline number.
• 9:09
  SPEAKER_00: Chinese firms shipped approximately 87% of all humanoid robots globally in 2025.
• 9:18
  SPEAKER_00: That's not a coincidence.
• 9:19
  SPEAKER_00: It's the result of a coordinated, state-directed industrial strategy.
• 9:24
  SPEAKER_00: China's Ministry of Industry and Information Technology set a target of over 100,000 humanoid robot units per year by 2026.
• 9:35
  SPEAKER_00: Year-over-year growth in 2025 was 594%.
• 9:41
  SPEAKER_00: To support this, China established the Beijing Humanoid Robot Innovation Center, released open source hardware specs, the Tiang Gong URDF files, and set up the TC08 Standardization Committee to define technical standards that could shape the industry for decades.
• 10:01
  SPEAKER_00: This is a playbook China has run before in solar panels, electric vehicles, and batteries, coordinate nationally, subsidize strategically, open source the platform, then dominate globally.
• 10:15
  SPEAKER_00: Now, what's the Western response?
• 10:17
  SPEAKER_00: It's primarily private, market-driven, decentralized, and in many ways extraordinarily innovative, but structurally very different.
• 10:26
  SPEAKER_00: Tesla's Optimus program is targeting a price point of$20,000 to$30,000 per unit, a potential democratization of humanoid robotics that$140,000 machines simply cannot achieve.
• 10:40
  SPEAKER_00: Boston Dynamics, now under Hyundai's ownership, has deployed its Atlas platform at the Hyundai MetaPlant in Georgia, scheduled for full integration by 2028.
• 10:51
  SPEAKER_00: Agility Robotics digit units are already operating inside Amazon fulfillment centers right now, and Unitary Robotics G1 platform is available between$16,000 and$90,000, making it one of the most price-competitive humanoid platforms in the world.
• 11:08
  SPEAKER_00: But here's the friction point.
• 11:10
  SPEAKER_00: The U.S.
• 11:11
  SPEAKER_00: Department of Defense has added Unitary Robotics to its restricted entity list.
• 11:16
  SPEAKER_00: The FCC has moved to ban certain Chinese manufactured communications hardware in American infrastructure, and export control frameworks on both sides are tightening rapidly.
• 11:26
  SPEAKER_00: What we are watching in real time is the bifurcation of the global physical AI supply chain into two increasingly incompatible ecosystems, Chinese platforms running on Chinese software stacks, and Western platforms operating under different regulatory frameworks, data governance rules, and security requirements.
• 11:48
  SPEAKER_00: If you're a manufacturer making decisions right now about which robotic platforms to invest in, you're being forced to choose a lane.
• 11:56
  SPEAKER_00: The hardware you buy this year will define your software dependencies for the next decade.
• 12:02
  SPEAKER_00: This is what analysts are calling the robot Cold War.
• 12:06
  SPEAKER_00: And unlike the original, which was about weapons and ideology, this one is about manufacturing capacity, about who physically makes the things the world uses every day.
• 12:17
  SPEAKER_00: What happens to global supply chains if 70% of humanoid robotics production becomes geopolitically inaccessible to Western firms?
• 12:26
  SPEAKER_00: These are not hypothetical questions for 2040.
• 12:29
  SPEAKER_00: They're being answered right now on factory floors and in government offices around the world.
• 12:35
  SPEAKER_00: Alright, let me pull the threads together and give you three concrete actionable takeaways.
• 12:41
  SPEAKER_00: Takeaway number one, stop watching the sprint times.
• 12:45
  SPEAKER_00: Watch the industrial task scores.
• 12:48
  SPEAKER_00: The performance benchmark that actually matters for the future of work is not how fast a robot can run 100 meters, it's how reliably, safely, and cost-effectively it can complete scenario-based industrial tasks, assembling components, navigating dynamic environments, interacting with human coworkers, recovering from unexpected situations.
• 13:11
  SPEAKER_00: The robots that crashed into the barriers in Beijing were optimized for one narrow metric.
• 13:18
  SPEAKER_00: The robots that will actually reshape manufacturing are the ones that score well on the 21 industrial task categories, the ones that handle the messy, unpredictable environments that real workplaces are.
• 13:31
  SPEAKER_00: If you're an investor, an engineer, a policymaker, or a business leader thinking about automation strategy, ignore the spectacle.
• 13:40
  SPEAKER_00: Follow the scenario scores.
• 13:42
  SPEAKER_00: That's where the signal is.
• 13:44
  SPEAKER_00: Takeaway number two, the supply chain bifurcation is not coming.
• 13:49
  SPEAKER_00: It's here.
• 13:50
  SPEAKER_00: And the window for building bridges is closing.
• 13:53
  SPEAKER_00: The regulatory divergence between the US and Chinese robotics ecosystems is accelerating.
• 14:00
  SPEAKER_00: DoD entity lists, FCC restrictions, export control frameworks, these reflect deep structural tensions likely to persist for years, possibly decades.
• 14:13
  SPEAKER_00: If you're an engineer, a policymaker, or an investor, you need to think about the geopolitical landscape your decisions are embedded in, which platforms to build expertise on, which companies can operate in a bifurcated world without being shut out of either market.
• 14:31
  SPEAKER_00: The organizations that will win are those that pick a lane decisively, or the rare, well-capitalized players that can genuinely operate in both ecosystems.
• 14:42
  SPEAKER_00: Those bridges are hard to build, but right now they're still buildable.
• 14:47
  SPEAKER_00: In five years, they may not be.
• 14:49
  SPEAKER_00: Takeaway number three, the next bottleneck is thermal, not computational.
• 14:56
  SPEAKER_00: We live in a moment where everyone is talking about AI, large language models, neural networks, and yes, that matters enormously, but the physical AI bottleneck over the next three to five years, it's heat.
• 15:09
  SPEAKER_00: Every time you push electrical current through a motor winding, you lose energy as heat.
• 15:15
  SPEAKER_00: Power lost equals current squared times resistance.
• 15:20
  SPEAKER_00: When a robot runs at peak sprint velocity or grips components repeatedly for hours, the currents are enormous.
• 15:28
  SPEAKER_00: The heat generated in compact joint housings, especially with QDD actuators in a humanoid form factor, is a serious engineering constraint.
• 15:38
  SPEAKER_00: We saw this in Beijing.
• 15:40
  SPEAKER_00: Thermal events contributed to post-race failures.
• 15:44
  SPEAKER_00: There's also the cascading mass penalty.
• 15:47
  SPEAKER_00: Add 200 grams at a robot's ankle for a better heatsink, and you need 350 grams at the knee, then 600 grams at the hip, then a larger battery.
• 15:58
  SPEAKER_00: Every design decision at the extremities propagates mass penalties up the entire kinematic chain.
• 16:06
  SPEAKER_00: The second engineering frontier is the sim-to-real gap.
• 16:10
  SPEAKER_00: Deep reinforcement learning trains robots in simulation, but the real world, with its real friction coefficients, floor textures, air resistance, and joint compliance variation, doesn't perfectly match the simulation.
• 16:25
  SPEAKER_00: Behaviors that work flawlessly in a virtual environment sometimes fail in unexpected ways on physical hardware.
• 16:32
  SPEAKER_00: Closing that gap is one of the hardest open problems in the field right now, and it's not a pure AI problem.
• 16:40
  SPEAKER_00: It's a physics modeling problem, a sensor calibration problem, a materials science problem.
• 16:46
  SPEAKER_00: The next wave of breakthroughs won't come from another transformer architecture or a bigger training run.
• 16:53
  SPEAKER_00: They'll come from materials scientists developing better thermal compounds, mechanical engineers designing smarter joint geometries, and simulation researchers building more faithful virtual environments.
• 17:06
  SPEAKER_00: The intelligence is largely there.
• 17:08
  SPEAKER_00: The body still has engineering problems to solve.
• 17:11
  SPEAKER_00: So let's take stock.
• 17:14
  SPEAKER_00: We started with a cold arresting fact.
• 17:16
  SPEAKER_00: A humanoid robot ran 100 meters faster than Usain Bolt.
• 17:21
  SPEAKER_00: We explored the architectural revolution from the tethered machines of the DARPA era to untethered sprinters at 14.5 meters per second, driven by deep reinforcement learning and quasi-direct drive actuators.
• 17:35
  SPEAKER_00: We examined the counterintuitive revelation, the winning robots crashed into the barriers, pointing to the deeper truth that sprint speed is not what matters.
• 17:45
  SPEAKER_00: Industrial task performance is.
• 17:48
  SPEAKER_00: We mapped the geopolitical chessboard, China's state-directed strategy producing 87% of global shipments, versus a Western model of private innovation, with the supply chain fracturing in real time.
• 18:02
  SPEAKER_00: And we landed on three takeaways.
• 18:04
  SPEAKER_00: Watch the task scores, not the sprint times.
• 18:07
  SPEAKER_00: The bifurcation is here and bridges are closing, and the next frontier is thermal engineering and simulation fidelity.
• 18:16
  SPEAKER_00: I want to close with something philosophical.
• 18:19
  SPEAKER_00: We have spent thousands of years defining ourselves, at least in part, by what our bodies can do.
• 18:26
  SPEAKER_00: We celebrated Hussein Bolt because he represented the outer limit of human physical possibility.
• 18:33
  SPEAKER_00: How fast can we go on two legs with training and will and biology?
• 18:39
  SPEAKER_00: And now a machine has exceeded that limit, not with arrogance, but simply because it was designed to, because engineers applied physics, materials science, and machine learning, and the result moves faster than any human body ever has.
• 18:56
  SPEAKER_00: I don't think that should make us afraid, but it should make us precise.
• 19:01
  SPEAKER_00: The question was never really how fast can we run.
• 19:05
  SPEAKER_00: The questions that define us have always been, what do we choose to build?
• 19:10
  SPEAKER_00: What do we use it for?
• 19:12
  SPEAKER_00: And who gets to decide?
• 19:14
  SPEAKER_00: Those questions are being answered right now in factories in Shenzhen, in labs in Ann Arbor and Boston and Seoul, in trade offices around the world.
• 19:25
  SPEAKER_00: The answers will shape the world your children inherit.
• 19:30
  SPEAKER_00: Thank you for spending this time with me on Mindcast.
• 19:33
  SPEAKER_00: If this episode gave you something to think about, a new frame, a question you hadn't considered before, I'd love to hear from you.
• 19:41
  SPEAKER_00: Leave us a review wherever you listen to podcasts.
• 19:44
  SPEAKER_00: It helps more people find the show.
• 19:46
  SPEAKER_00: All the stats, researchers, and concepts from today's episode are in the show notes, with links to further reading on deep reinforcement learning, the geopolitics of physical AI, and the engineering bottlenecks we discussed.
• 20:00
  SPEAKER_00: Subscribe to Mindcast so you never miss an episode.
• 20:03
  SPEAKER_00: I'm Will.
• 20:04
  SPEAKER_00: This has been Mindcast.