Latest / Elon Musk Podcast / Mcdonald's Hires AI Robot in China

Transcript

Mcdonald's Hires AI Robot in China · Elon Musk Podcast

• 0:00
  McDonald's has actually placed a bipedal humanoid robot built by
• 0:03
  Keenan Robotics right on a restaurant floor in Shanghai to
• 0:07
  greet customers like it's wearing a custom red and yellow
• 0:09
  uniform and everything. Yeah.
• 0:10
  Seeing a walking, talking machine navigating A crowded
• 0:14
  fast food space is just startling.
• 0:17
  It really is. You are looking at a live
• 0:19
  deployment of embodied artificial intelligence
• 0:22
  interacting directly with the public.
• 0:24
  I mean we are looking at the exact point where factory
• 0:26
  automation bridges the gap and walks right into everyday
• 0:29
  customer. So if a global burger chain is
• 0:33
  successfully putting bipedal robots in front of customers
• 0:36
  today, what happens to the millions of human jobs that rely
• 0:40
  on the service industry tomorrow?
• 0:41
  Well, the machine operating inside that Shanghai Science and
• 0:44
  Technology Museum, McDonald's is specifically the Keenan X Man
• 0:48
  F1, and to understand why this is working, you really have to
• 0:51
  look at the physical construction.
• 0:52
  Right, the hardware. Exactly.
• 0:54
  It features 43° of freedom, so for context, a standard
• 0:59
  industrial robotic arm might have like 6° of freedom.
• 1:04
  Just enough to move up, down, left, right, rotate.
• 1:06
  Yeah, exactly. But 43° means the machine has
• 1:09
  individual articulation in its neck, shoulders, elbows, wrists,
• 1:13
  and even its fingers. It allows for incredibly fluid,
• 1:18
  highly nuanced physical movement.
• 1:20
  Movement that actually mimics a human being navigating physical
• 1:23
  space and driving. All of that articulation is a
• 1:28
  vision Language action model. Specifically their KOM 2 point O
• 1:32
  system. Right, because a standard
• 1:33
  language model just processes text.
• 1:35
  Yeah, but a vision Language action model does exactly what
• 1:38
  the name implies. It sees the physical world
• 1:40
  through cameras, processes, spoken language from a customer,
• 1:43
  and instantly translates both of those inputs into a physical
• 1:47
  action. Real time, and the most crucial
• 1:49
  part of how this operates on a restaurant floor is a concept
• 1:52
  Keenan calls pro S vocational training.
• 1:54
  Pro S. Yeah, they're not just giving
• 1:57
  the robot a general intelligence and hoping it figures out how to
• 1:59
  work in McDonald's. This software approach
• 2:01
  deconstructs highly complex service scenarios into
• 2:04
  independent, completely standardized work modules.
• 2:07
  So it operates almost like downloading a highly specific
• 2:10
  job description right into a physical body?
• 2:12
  Pretty much, yeah. Like, rather than programming a
• 2:14
  machine from scratch to just wander around a room, you are
• 2:18
  handing it a digital binder. You tell the system you know you
• 2:22
  are the greeter. Here are your specific
• 2:25
  parameters for exactly how to behave when the door opens or.
• 2:29
  You are the demonstrator. Here is the exact routine you
• 2:32
  run when a customer asks about a menu item.
• 2:35
  And the public response in Shanghai reflects how effective
• 2:38
  that programming really is. Oh, the robot was a massive hit
• 2:42
  with families eating there because of those 43° of freedom.
• 2:46
  It was capable of making heart shapes with its hands.
• 2:49
  Oh, wow. Yeah.
• 2:50
  It interacted directly with kids responding to their movements.
• 2:53
  It was actually so well received that customers gave it the
• 2:56
  nickname Alphabot. That shifts the robot from being
• 2:59
  just a generalized tool into a specialized employee.
• 3:02
  I mean, it opens up the ability for corporations to enforce
• 3:05
  absolute standardization in physical customer service.
• 3:08
  Right, when a company deploys a system like this, it ensures
• 3:12
  every single interaction is identical.
• 3:14
  Every customer gets the exact same greeting, the exact same
• 3:17
  posture, the exact same level of enthusiasm.
• 3:20
  It completely removes the variable of human mood or
• 3:23
  fatigue. And that standardization
• 3:25
  addresses a massive financial leak in the service sector.
• 3:29
  The fast food industry currently faces severe turnover, reaching
• 3:33
  a staggering 150% annually. A. 150%.
• 3:37
  Yeah, you have restaurants completely replacing their
• 3:40
  entire staff and then half of them again every single year.
• 3:44
  Research shows that costs businesses approximately $1600
• 3:48
  per lost employee. Because of the time spent
• 3:50
  interviewing, the cost of the uniform, the training shifts.
• 3:54
  Exactly, plus the inevitable mistakes made along the way when
• 3:57
  they're new. Meanwhile, McDonald's is pushing
• 3:59
  to expand to 50,000 global locations.
• 4:03
  The logistical nightmare of finding, hiring, and retaining
• 4:06
  enough humans for that with a 150% turnover rate is almost
• 4:11
  insurmountable. Right.
• 4:12
  And for context on how automation solves this
• 4:14
  mathematically, just look at competitor Yum China.
• 4:17
  What did they do? They managed to grow their store
• 4:19
  count by 56%. They added thousands of physical
• 4:24
  restaurant locations to their portfolio, but they kept their
• 4:27
  human workforce entirely flat at 420,000 people.
• 4:31
  Wow. They achieved that growth purely
• 4:34
  through heavy automation. And Keenan's explicit business
• 4:37
  model fits right into that exact math.
• 4:40
  They actually refer to themselves as a labor
• 4:42
  outsourcing company. Yeah, they're not just selling a
• 4:45
  piece of hardware. No, they're leasing these robots
• 4:48
  for less than half the cost of human labor.
• 4:51
  You are basically renting an employee.
• 4:53
  Think about growing a business by half without hiring a single
• 4:56
  new person. That is, yeah, yeah.
• 4:58
  That completely limits the need for future human hiring in high
• 5:02
  growth markets. It changes the financial math of
• 5:04
  expansion entirely. It makes robots a required
• 5:07
  baseline rather than a futuristic luxury.
• 5:10
  Right. You simply cannot compete if
• 5:12
  your labor costs are double your competitors because you were
• 5:15
  paying human wages and absorbing that $1600 turnover penalty.
• 5:19
  Not while the restaurant across the street is leasing a machine
• 5:22
  for a fraction of the price. But the X MNF 1 humanoid doesn't
• 5:26
  actually work alone in these environments.
• 5:28
  It relies on a matrix of other robots.
• 5:31
  OK, a matrix. Yeah, the system coordinates
• 5:34
  directly with Keenan's Diner Bot T series.
• 5:37
  Those are wheeled robots completely lacking A humanoid
• 5:40
  shape, and they handle the heavy lifting.
• 5:42
  Transporting food trays from the kitchen out to the dining room
• 5:44
  floor, right? Exactly.
• 5:46
  In some specific hardware configurations, Keenan even
• 5:50
  swaps the humanoids legs for specialized attachments so the
• 5:53
  machine can perform highly specific tasks faster than a
• 5:57
  bipedal walking motion allows. Wait, backup.
• 5:59
  If they remove the legs, doesn't that defeat the whole purpose of
• 6:03
  having a bipedal humanoid on the floor?
• 6:05
  You would assume so, yeah. But bipedal movement is really
• 6:07
  just one tool in the box. Walking on 2 legs is
• 6:10
  computationally expensive and relatively slow when you're
• 6:14
  carrying heavyweight. So why have the humanoid at all?
• 6:17
  Because the humanoid serves as the social, relatable interface
• 6:20
  for the customer, it provides the smile, the eye contact, the
• 6:23
  verbal interaction. While the wheeled robots operate
• 6:26
  as the highly efficient unseen logistics network, moving the
• 6:29
  actual product around. Exactly.
• 6:31
  You have a friendly face taking your order while a highly
• 6:34
  efficient rolling cart handles the actual delivery of the the
• 6:37
  tray. That changes the vision of
• 6:39
  automation from a single stand alone machine doing everything
• 6:43
  to a coordinated swarm of specialized agents working
• 6:46
  together. And for humanoids to succeed as
• 6:49
  that social interface, they must be able to mimic human
• 6:53
  connection convincingly. You see this heavily prioritized
• 6:57
  in models like the Ahead Form desktop robot.
• 7:00
  This machine achieves human like connection using tiny brushless
• 7:04
  motors placed precisely under synthetic skin.
• 7:07
  To create micro expressions. Yeah, these motors pull and
• 7:10
  release the material so the robot blinks, softens its eyes,
• 7:14
  and smiles in perfect synchronization with its speech
• 7:17
  output. It even implements a short pause
• 7:19
  and a downward gaze before answering a question to simulate
• 7:22
  human thought. Which is wild.
• 7:24
  It is that slight pause is just like a conversational breath,
• 7:27
  like a response that comes out too fast with 0 hesitation feels
• 7:31
  completely robotic and unnatural.
• 7:33
  But programming a one second delay accompanied by a physical
• 7:37
  hift, an eye contact away from the user and then back, feels
• 7:41
  remarkably human. It signals to your brain that
• 7:45
  the entity is processing what you just said.
• 7:47
  Because people respond fundamentally differently to a
• 7:50
  physical machine that makes eye contact versus a Voice Assistant
• 7:54
  on a phone. Definitely.
• 7:55
  When a physical entity occupies your 3 dimensional space, turns
• 7:59
  its head to look at you, and physically reacts to your facial
• 8:03
  expressions, your brain categorizes it differently than
• 8:07
  a smart speaker sitting on a kitchen counter.
• 8:09
  And this opens up robotics to handle emotional labor.
• 8:12
  It changes the role of machines from purely physical workers
• 8:15
  moving boxes to entities that can address loneliness, provide
• 8:19
  companionship and elevate customer satisfaction.
• 8:22
  Yeah, a customer feels heard and respected when the machine
• 8:24
  taking their order softens its eyes and smiles at them.
• 8:27
  But operating heavy bipedal robots near humans does carry
• 8:30
  inherent physical risks. Oh, absolutely.
• 8:32
  A viral video captured a unitary H1 humanoid robot suffering a
• 8:36
  severe malfunction while suspended from a crane in a
• 8:38
  factory setting. It completely lost control of
• 8:42
  its motor functions. It flailed its arms and legs
• 8:45
  wildly and dragged its metal stand across the the floor until
• 8:48
  it smashed a computer. Wait.
• 8:50
  Hold on, it just started swinging its arms and destroyed
• 8:53
  a computer. Yes, the erratic movements were
• 8:55
  intense and highly destructive, and a previous unitary incident
• 8:59
  even involved a robot lunging unexpectedly toward a crowd of
• 9:02
  people. That's terrifying.
• 9:04
  When a machine weighing over 100 lbs loses spatial awareness, it
• 9:08
  becomes a severe hazard. But Keenan attemts to mitigate
• 9:12
  these exact risks in their models using reinforcement
• 9:15
  learning based dynamic balance. Like running simulations.
• 9:18
  Millions of simulations so the robot learns exactly how to
• 9:21
  correct its posture if it gets bumped.
• 9:23
  They also equip the robots with 3D LIDAR, which bounces light
• 9:27
  around the room to build a real time topographical map of the
• 9:31
  space. Along with force feedback
• 9:32
  sensors. Right.
• 9:33
  Those sensors allow the robot to feel physical resistance,
• 9:36
  ensuring it can safely navigate tight, crowded restaurant aisles
• 9:39
  without forcefully colliding with customers.
• 9:42
  These incidents limit how quickly high mass bipedal units
• 9:45
  can be deployed in crowded public spaces, though it creates
• 9:48
  an entirely new category of physical liability for
• 9:51
  businesses adopting embodied AI. For sure.
• 9:54
  I mean, a software glitch in a customer service chat bot gives
• 9:57
  you a weird sentence on a screen.
• 9:59
  A software glitch in a highly articulated heavy metal biped
• 10:03
  can cause serious physical damage to a customer or a
• 10:06
  storefront. And to prevent those errors and
• 10:09
  allow robots to adapt safely to unpredictable environments,
• 10:13
  manufacturers are radically upgrading the software brains
• 10:16
  powering the hardware. Boston Dynamics partnered with
• 10:19
  Google DeepMind to integrate the Gemini AI model into their
• 10:23
  Electric Atlas robot, didn't they?
• 10:26
  Yes, they did. So instead of being rigidly pre
• 10:28
  programmed to walk a specific path, Atlas can now look at a
• 10:31
  room, see a box sitting in the wrong place, understand it needs
• 10:34
  to be moved to a specific bin, and plan The Walking path
• 10:37
  entirely on its own. Using real time spatial
• 10:40
  reasoning? Exactly.
• 10:41
  I have to say, giving a highly strong physical robot the
• 10:43
  ability to reason and make its own decisions sounds incredibly
• 10:46
  dangerous. Really.
• 10:48
  Allowing a machine of that size to improvise its movements
• 10:51
  introduces a massive variable into a public space.
• 10:55
  If it decides on its own how to get from point A to point B, you
• 10:59
  lose the predictability of a programmed machine.
• 11:02
  I see what you're saying, but reasoning is exactly what
• 11:05
  prevents the robot from blindly executing a task and walking
• 11:08
  into a wall or a person. How so?
• 11:10
  If a pre programmed robot is told to walk exactly 10 feet
• 11:14
  forward and a child steps in the way at foot 5, the robot walks
• 11:18
  right into the child because it has no concept of the
• 11:20
  environment. A reasoning robot understands
• 11:23
  the environment has changed, recognizes the new obstacle, and
• 11:27
  instantly stops or alters its path.
• 11:29
  OK, that's a fairpoint, and Boston Dynamics recently
• 11:32
  released an unedited video of the research Atlas failing,
• 11:36
  slipping on ice, and completely collapsing during backflip
• 11:39
  attempts. Yeah, the machine miscalculates
• 11:40
  its landing, its joints give out and it crashes into the mat.
• 11:44
  It proves that the robots learn through making mistakes in
• 11:46
  complex environments. They're physically testing the
• 11:49
  limits of their own hardware, building an inherent
• 11:51
  understanding of gravity, momentum and friction.
• 11:54
  Which changes robots from rigid pre programmed tools into
• 11:58
  adaptable workers capable of learning entirely new tasks in
• 12:02
  less than a day based purely on observation.
• 12:05
  They're developing A fundamental understanding of physics and
• 12:08
  spatial relationships, allowing them to operate safely in spaces
• 12:11
  built for humans. Because attempting to automate
• 12:14
  physical spaces like fast food with software alone has distinct
• 12:18
  limits, McDonald's recently had to end its AI drive through
• 12:22
  trial with IBM. Oh yeah, they attempted to use
• 12:25
  purely voice based artificial intelligence to take customer
• 12:28
  orders and the system went viral for severe misinterpretations.
• 12:32
  It was adding incorrect items like single butter packets to
• 12:35
  orders. Or ringing up hundreds of
• 12:36
  dollars worth of Mcnuggets as frustrating customers laughed
• 12:39
  and repeatedly begged the microphone to stop adding items
• 12:42
  to the screen. Right, because software only AI
• 12:44
  struggles with the unpredictable nuances of human interaction in
• 12:47
  the physical world. Accents.
• 12:49
  Loud background noise from an engine.
• 12:51
  Hesitations. Yeah, or a customer changing
• 12:54
  their mind mid sentence. All of that completely derails a
• 12:57
  system that lacks physical context.
• 13:00
  The software only hears the audio, it cannot see the
• 13:03
  environment. Which is exactly why embodied AI
• 13:06
  machines equipped with cameras, spatial awareness, and a
• 13:08
  physical presence is necessary to bridge the gap that chat bots
• 13:12
  cannot cross. A robot standing on the floor
• 13:15
  can see if a customer is physically pointing at a
• 13:17
  specific menu item or if they look visually confused.
• 13:20
  The machine uses the visual data to add layers of context to the
• 13:23
  verbal command, allowing it to understand what the human
• 13:26
  actually wants. This limits the effectiveness of
• 13:29
  invisible software, only automation.
• 13:32
  It proves that to truly automate physical spaces, businesses
• 13:36
  require physical machines that can interpret physical context.
• 13:39
  You cannot run a restaurant with just a microphone and an
• 13:41
  algorithm. You need hardware that can
• 13:44
  observe the room. Yeah, absolutely.
• 13:46
  The deployment of humanoid robots in places like McDonald's
• 13:49
  shows that automation is stepping out from behind the
• 13:51
  counter and into the physical world.
• 13:54
  It merges social interaction with calculated efficiency,
• 13:57
  putting a red and yellow uniform on a highly advanced machine to
• 14:01
  provide a standardized, scalable workforce.
• 14:04
  But when the friendly, smiling robot handing you a burger costs
• 14:07
  half as much as a human, how long before human interaction
• 14:10
  becomes a premium luxury you have to pay extra for?
• 14:12
  If you're not subscribed yet, take a second and hit follow on
• 14:15
  whatever app you're using. It helps us keep making this.
• 14:17
  We appreciate you being here.