Specialized robots are winning first, from warehouses to brain surgery

The Short Version

The robots that are earning their place today are built around one job, and they get that job because they do it better than a person can. Robert Playter, the CEO of Boston Dynamics, says the accumulated wisdom of the field is to "go build a robot for a specific task and it'll do it very well," and he dropped his company's most graceful machine, Handle, for a plainer one, Stretch, because Stretch unloads trucks faster [1]. Matthew MacDougall, Neuralink's head neurosurgeon, explains that a robot places the company's electrodes because no human hand is steady or fast enough, while a human surgeon still does everything else in the operation [2]. Kate Darling, who studies human-robot interaction at the MIT Media Lab, argues that the humanoid form is "a distraction" and that robots, like working animals, are worth having when their skills are different from ours [3]. None of the three says general-purpose robots are pointless. Playter is still building one, and he is clear that it is the harder road and the later one [1].

Table of contents

What They Said

Robert Playter: the beautiful robot that lost to the useful one

Playter has been with Boston Dynamics since its roots in Mark Raibert's Leg Lab at MIT, and his clearest lesson about general-purpose machines came from a contest his company supplied but did not enter [1]. For the 2015 DARPA Robotics Challenge, Boston Dynamics built about ten first-generation Atlas humanoids and handed them, with baseline software, to competing teams [1]. The tasks were ones a person would not think twice about: drive a vehicle, get out, open a door, close a valve, cut a hole with a tool, and cross stairs and rough terrain, all with weak and sometimes interrupted communication [1]. Playter's summary of the result is blunt. Mostly the robots failed and fell down, and he calls the event humbling [1]. "A person could have done it very easily," he says of the tasks [1].

He draws two points out of that event. The first is what he calls the curse of a general-purpose robot: it is not perfect at any one thing. Atlas was a little too big to get in and out of the car, which is the kind of problem a machine designed only for that car would never have [1]. The trade is that a general robot can do many things and cope with uncertainty [1]. The second is how much hidden difficulty sits inside an ordinary action. He recalls a robot that reached for a door handle, missed the grasp, and then threw itself over when it tried to turn the knob, because it expected the door to push back and nothing did [1]. In his words, "the first time you start to try to push on the world with a robot, you, you forget that the world pushes back" [1].

Playter states the contrast directly. "I think we all want to build general-purpose robots that can be used for lots of different activities. But it's hard," he says, and then gives the field's working rule: "the wisdom in building successful robots up until this point have been go build a robot for a specific task and it'll do it very well. And as long as you control that environment, it'll operate perfectly" [1]. He does not treat that as the end of the story. Robots that are going to be useful in the future have to deal with unexpected situations, he says, and that is the goal of a general-purpose or multi-purpose robot. "And that's just darn hard" [1].

The decision that best shows where he comes down is Handle. Handle grew out of the Atlas balance research: once the team understood how to control momentum in a body with heavy mass up high, it built a two-wheeled balancing robot with legs, a long arm to reach boxes and a counterbalancing tail, all moving together [1]. Playter calls it graceful like nothing else the company has built [1]. It was also too slow at picking boxes for logistics, so they let it go, a choice he describes as being about building a business rather than chasing beautiful mathematics [1]. He thinks the company will come back to it eventually [1].

Its replacement is deliberately less clever. Stretch, he says, "looks like a big, strong robot arm on a mobile base," and the base is about the size of a pallet, because pallets are what live in warehouses and the robot has to fit in the same space [1]. It does not balance at all. Playter says it was "a bit of commitment from us, a challenge for us to build a non-balancing robot," which for a company famous for legged machines meant choosing the easier engineering problem on purpose [1]. The base is omnidirectional, and its big battery gives about 16 hours of operation, or two shifts, and helps the robot handle boxes of up to 50 pounds without tipping [1].

The job it was built for is one people do badly for reasons that have nothing to do with skill. An estimated trillion cardboard boxes are shipped each year, and trucks are unloaded by hand inside containers that can reach 120 degrees [1]. Customers wanted Stretch immediately. Deliveries began in January to DHL and Maersk, Gap is next, and commitments cover a couple hundred robots [1]. Playter says the workers who used to unload the trucks are now operating the robots, and customers typically see a return on their investment in about two years [1].

Spot, the yellow robot dog, looks like a general machine but sells as a specialist. Its main market is industrial inspection: thermal cameras to catch overheating, acoustic imagers to find compressed-air leaks, cameras to read gauges, and vibration sensing, in places such as AB InBev breweries, GlobalFoundries chip fabs, utilities and nuclear plants [1]. The economics come from what a missed fault costs. A chip fab line that goes down because a vacuum pump failed can cost a million dollars a day [1]. For Ontario Power Generation, Spot operates a dangerous high-voltage breaker switch where arc flashes can kill people, and the company automated that single job so that an operator just marks the switch and socket on a tablet and the robot does the rest [1]. The next software release, version 3.3, adds one more narrow skill: opening a door on its own during a mission, after an operator points out the handle and says which side the hinges are on [1].

Behind all of this is arithmetic Playter does not soften. "If you only sell hundreds, you will commercially fail," he says. A robot company needs to sell thousands or tens of thousands, and that takes patient investors [1]. About 1,100 Spots are in the field, and he says that part of the business needs perhaps 1,000 to 1,500 a year to be profitable [1]. Of Stretch he says, "There's definitely thousands of robots," and probably tens of thousands in the future [1]. The same arithmetic explains why the company builds for industry and not the home. When Google owned Boston Dynamics, Larry Page kept asking what the product would be and favored everyday consumer products costing a few thousand dollars, which Playter calls the toothbrush test. Boston Dynamics was building machines costing hundreds of thousands of dollars or more, and industry is the customer that can afford a robot at a hundred thousand dollars and up [1].

He applies the same rule to the home robot that Lex Fridman keeps asking about. The path there, Playter says, runs through industrial applications and then mass manufacturing. "The principle of the company has always been really make the robots do useful stuff," he says, and the social robot companies that started with a cute interaction mostly have not survived. "And so we think the utility really needs to come first" [1]. A home robot would need real safety around children and the ability to handle varied tasks, such as telling a beer from a cola or clearing dishes without breaking them [1].

Playter has not given up on the general machine. Natural walking on Atlas took about 15 years to arrive, roughly 40 developers share just four modern Atlas robots, and the next steps he names are more dexterous hands and moving heavy, unwieldy objects with two arms in factories or construction, as opposed to fine work [1]. He even expects the specialist to widen: Stretch will eventually do any box-moving task in the warehouse, including palletizing, depalletizing and loading trucks as well as unloading them. "And so it, too, will be a multi-purpose robot," he says [1]. He sees mobile manipulation as the hard problem of the next ten years [1].

Matthew MacDougall: a robot for one step of the operation

MacDougall's robot is narrower still. What Neuralink is building, he says, is a neural implant plus a robot that inserts electrode threads thinner than a human hair into a small region of the brain [2]. The first people it is meant for have severe spinal cord injuries and are quadriplegic. Their motor cortex is intact but disconnected from their limbs, and the implant would read their movement intentions so they can control a computer mouse and keyboard, which he calls digital freedom [2].

The reason for the robot is physical. The electrodes are so tiny, and the blood vessels on the brain's surface so dense, that no human hand is steady or fast enough to place them accurately at the right depth [2]. This is the plainest version of the pattern in any of the three conversations: the machine is there because the task is beyond what a person's body can do, not because it is cheaper or more convenient [2].

The second half of his explanation matters as much. A neurosurgeon still opens the skin and the skull. The robot handles only the electrode placement [2]. Nothing about the judgment-heavy parts of the surgery has been handed over. One step, the one that exceeds human steadiness and speed, has been given to a machine built for exactly that step [2].

MacDougall is also specific about how limited the system is. Current technology can reach only the brain's surface, so the deep functions people most hope to treat, such as mood, appetite, addiction, pain and sleep, will come later [2]. At this stage patients will not move their own limbs. Imagining a movement moves a cursor, opens a word processor and types on an on-screen keyboard [2]. He says he would have the implant himself, but that it would not be useful to him yet, because early devices do things he can already do with his hands [2]. The narrow first application was chosen because it is where the device clearly beats the alternative, which for a quadriplegic patient is no independent control of a computer at all [2].

He backs the approach with a safety claim from his own experience. From the hundreds of surgeries he has performed with it, he says the procedure is already safer than many standard FDA-approved operations [2]. The source does not give figures behind that statement, and at the time of the conversation human trials were still being finalized with the FDA, with a patient registry open for people with quadriplegia [2]. He also notes that pigs serve as the safety platform because their skulls are similar in size and shape to human skulls, while monkeys provide the functional data [2]. Anyone weighing a brain implant or any neurosurgical procedure should do so with their own neurosurgeon and medical team.

MacDougall's wider comments show the same instinct for matching a tool to the job it can actually do. Asked about boosting brain plasticity in adults, the Neuralink surgeon says drugs will be more fruitful than implants, because changing plasticity means altering something like a trillion synapses across the whole brain, which a drug can reach and an electrode cannot [2]. He also describes laser interstitial thermal therapy, in which a fiber-optic probe is passed through a two-millimeter hole to heat a deep tumor while an MRI scanner monitors the temperature in real time, killing tumor cells while sparing the tissue around them [2]. Tumors once considered inoperable can sometimes be removed this way with patients coming out essentially unscathed [2]. In each case a tool built for one precise act does what general skill could not.

Kate Darling: stop building copies of ourselves

Darling comes at the question from the other side. She does not build robots; she studies how people respond to them, and her book argues that we should stop comparing robots to humans and look instead at our long history with animals [3]. Her target is the picture in most people's heads. Search for images of a robot, she says, and you get humanoids with a torso, a head, two arms and two legs, and that constant comparison of robots to people, and of artificial intelligence to human intelligence, trips us up [3]. AI does not think like people, and she finds it boring to recreate an intelligence we already have when we could create new skill sets to partner with [3].

She takes the two standard arguments for humanoids in turn. The first is that people relate best to things that look like themselves. Social robotics research, she says, shows you only need enough cues for people to relate, and the thing giving off the cues does not have to look or act human. Her example is R2-D2, which just beeps and is, as she puts it, "just like a trash can on wheels," yet people like it more than the humanoid C-3PO [3]. The second argument is that the world is built for human bodies, with stairs, narrow passageways and door handles [3]. She grants that a few things cannot be re-engineered, such as a submarine or a spaceship. But a factory is different: "it seems like we've already rebuilt factories to accommodate other types of robots," she says. "Why would we want to just like make a humanoid robot to go in there?" [3]

When Fridman objects that most of the world is still built for human bodies, she answers that it should not be: "It should be built so that it's wheelchair accessible" [3]. She would put investment into infrastructure that works for, in her words, "all kinds of people, all kinds of robots," and she asks why a robot should not have three arms or roller skates [3]. She is careful not to dismiss humanoids completely. She respects them as a research problem and as a magnet for talent, says it does not have to be either or, and thinks people are fascinated by them. But she calls the form "a distraction" for social robotics and "one that I find particularly boring," and says of the excitement, "I think it's a little overhyped" [3].

Her positive case is the animal analogy. We domesticated animals not because they do what we do, but because what they do is different and useful [3]. The ox transformed agriculture and whole economies. Ferrets run wires through narrow pipes. Pigeons carried mail and were early aerial photographers. The U.S. and Russian navies still use dolphins for mine detection because their echolocation beats machines [3]. Each of those is a specialist with one ability a person lacks. Robots, she argues, will reshape economies and even city design in the same way, and the analogy, which she admits is not perfect, gets us past the idea that robots will simply replace people [3].

Her read of the present matches Playter's. "We're way behind where people think we are," she says of physical robotics [3]. The milestone that impresses her is Amazon's first autonomous warehouse robot that is safe for people to be around, because strong industrial robots have usually had to be separated from people or confined to fixed lines [3]. It is very difficult, she says, to know where the human is and to respond dynamically, especially for something as powerful as an industrial arm [3]. On jobs her answer is "I don't think it's that simple" [3]. Warehouse work where people are unsafe and badly treated might be better done by a robot, and in mining, autonomous trucks let workers sit in air-conditioned control stations instead of at dangerous sites, though fewer people are employed [3]. Over time more jobs are created, but it is easy to ignore what she calls "the rubble on the ground" during the transition [3].

She also supplies the limit on the pattern. Self-driving cars are taking longer than expected because of the long tail of unexpected situations, like a horse and carriage on the highway, and she thinks the key is combining human and machine skills and solving the handoffs between them [3]. Fridman adds an example from Ukraine: demining is still mostly done by humans and animals, despite the obvious appeal of robots, because disarming mines takes enormous experience with edge cases [3].

Fridman is the dissenting voice in her conversation. He describes Tesla's Optimus humanoid, meant for boring factory tasks, and says he has learned not to discount Elon Musk's efforts [3]. What excites him is the first-principles question of cost: humanoids have been expensive, he argues, because their builders were not focused on mass manufacture, and he sees no good reason one could not be built far more cheaply [3]. He also argues that whoever solves humanoid robots will have to solve social robotics along the way [3].

Where They Agree, and Where They Don't

All three agree on where things stand. Playter says the robots that have succeeded so far were built for a specific task in a controlled environment [1]. MacDougall's surgical robot does one step of one operation [2]. Darling says physical robotics is far behind what people imagine and that the impressive recent step was a warehouse robot that can safely share space with workers [3].

They also agree that the winning robot is doing something a person does poorly or cannot do. For MacDougall it is placing hair-thin threads between dense blood vessels [2]. For Playter it is lifting 50-pound boxes for 16 hours in a 120-degree container, or working a breaker switch that can kill the person standing at it [1]. For Darling it is the principle behind every working animal: the value lies in the ability that differs from ours [3]. And they agree that the human stays in the picture. The neurosurgeon opens the skull [2], the truck unloaders become robot operators [1], and Darling describes robots and people as having very different skill sets that work best together [3].

The disagreement is about the humanoid. Darling finds it a boring distraction and would redesign the environment for all kinds of bodies [3]. Playter has spent decades on Atlas, says everyone wants general-purpose robots, and argues that dealing with the unexpected is what will make robots useful in the future [1]. The gap is narrower than it looks. Playter's own account supports much of her skepticism: he calls generality a curse, admits the specific-task robot is the proven recipe, and chose a pallet-sized arm on wheels over a balancing machine when a business depended on it [1]. And Darling concedes that humanoids are a worthwhile research problem [3]. Where they truly part is on what the humanoid is for. Playter treats it as a long-term goal and a source of knowledge that feeds his products, since Handle came straight out of Atlas balance research [1]. Darling asks why anyone would aim there at all when a factory can be rebuilt [3].

They differ on the social side too. Playter says social robot companies that put a cute interaction first mostly have not survived, and that usefulness has to come first [1]. Darling shows that usefulness alone is not enough once a robot shares space with the public. Marty, the tall robot that looks for spills at Stop & Shop, has a single practical job, yet a survey found people complained that it felt like a creepy surveillance robot that got in the way, and the biggest spike in negative social media mentions came when the store threw the robots a birthday party [3]. She suspects the big googly eyes made people feel watched, and wonders whether Marty is "a real life clippy" [3]. The two views fit together more than they clash. Playter himself says body language will matter in practice, because industrial operators need to see what a robot is doing, whether it needs help, or that it is about to turn [1].

Why It Works

The sources give four reasons that the narrow robot wins first.

A controlled environment removes most of the surprises. Playter's formula has two halves: a specific task, and an environment you control [1]. The DARPA robots fell because the world kept doing things they had not predicted, like a door that did not push back [1]. Coping with that takes layered control loops, running at roughly a thousand hertz at each joint and about a hundred hertz for whole-body motion, plus slower loops for perception, with constant prediction and fast re-planning [1]. A warehouse built around pallets, or a patch of brain surface already exposed by a surgeon, takes much of that uncertainty away [1][2].

A narrower machine is a simpler machine. Stretch does not balance, which Playter describes as an easier problem than the ones his company usually takes on [1]. A humanoid, by contrast, is underactuated, since the only outside forces come through its feet, and its heavy upper body and heavy legs make balance much harder than on a four-legged robot [1]. Simpler machines are also easier to make reliable. To find rare failures on Spot, a fleet runs autonomous missions around the office, logging something like 1,500 to 2,000 kilometers and over a thousand hours a week [1].

The task is judged against a person, and the machine clearly wins. No hand can place Neuralink's threads [2]. No person should work an arc-flash breaker or unload trucks in that heat [1]. When Handle did not beat the alternative on box-picking speed, it was dropped despite its elegance [1]. Physical work also makes the comparison honest. Fridman tells Playter that "The physical reality is a good verifier," and Playter agrees: with a manual task you can easily judge whether the robot did it and did it correctly [1].

One clear job makes the money work. A two-year payback sells Stretch [1]. A million dollars a day of downtime sells inspection by Spot [1]. Sales in the thousands are what keep a robot company alive, and a buyer with one expensive, well-defined problem is the buyer who can pay for a machine that costs a hundred thousand dollars or more [1].

Where It Applies, and Where It Doesn't

The pattern holds best where the task repeats, the setting can be arranged around the machine, and a person is either physically unable to do the job well or is put in danger by it. Truck unloading, gauge reading in a chip fab, breaker switching and electrode insertion all fit [1][2]. So do Darling's examples of mining trucks and warehouse work [3].

It weakens where the edge cases never end. Demining and highway driving are both narrow jobs on paper, and both have resisted full automation because of the long tail of situations nobody planned for [3]. Darling's answer there is a human and a machine working together, with the handoff between them treated as the real problem to solve [3].

It also does not settle the long-term question. Playter expects the specialists to broaden, with Stretch taking on more warehouse tasks and Spot gaining an arm that opens doors, and he still sees the general robot that copes with uncertainty as the goal [1]. The home is where the narrow approach runs out, because a household asks for many varied tasks, done safely around children [1]. And for some purposes a robot is the wrong tool entirely. MacDougall would use a drug, not an electrode, to change plasticity across the whole brain [2].

Finally, a robot that does its one task well can still fail with the people around it, as Marty shows [3]. Doing the task better than a person is necessary. In public spaces it is not sufficient.

What To Do With It

1. Judge any robot by one question: what single task does it do better than a person? Stretch has an answer in boxes, hours and degrees of heat, and Neuralink's robot has one in the width of a thread [1][2]. If the pitch is that a machine can do a bit of everything, remember Playter's own verdict that such a robot is not perfect at any one thing [1].

2. Start with the dangerous, punishing or physically impossible job. The clearest wins in these conversations are a 120-degree truck container, an arc-flash breaker and threads no hand can place [1][2]. Darling makes the same case for warehouse work where people are unsafe [3].

3. Be willing to drop the impressive option for the efficient one. Playter let go of the most graceful robot his company had built because it picked boxes too slowly, and built a non-balancing arm on a pallet-sized base instead [1].

4. Ask whether you can change the environment before you ask for a humanoid. Darling's point is that factories have already been rebuilt for other kinds of robots, and that spaces designed to be accessible work for all kinds of people and all kinds of robots [3].

5. Give the robot one step and keep the person for the rest. The neurosurgeon opens the skin and skull and the robot only places the electrodes [2]. With Spot, an operator marks the switch and socket on a tablet and the robot does the dangerous part [1].

6. Run the numbers before the demo. Playter's measures are payback in about two years, the daily cost of downtime, and whether the product can sell in the thousands [1].

7. Test for the rare failure, not the average run. Boston Dynamics keeps a fleet of Spots walking missions around its office for more than a thousand hours a week to surface the failures that only show up at scale [1].

8. Plan for how people will react. Give a robot a name, since naming helps people accept it and forgive its mistakes, but avoid gendering it according to stereotypes about its function, and make sure people can read what it is doing [3][1]. Marty's googly eyes and birthday party show what happens when that is left to chance [3].

9. Treat the people whose work changes as part of the plan. At Stretch's first customers the truck unloaders became the robot operators [1]. Darling's warning is that the long-run gains do not erase the damage to the people caught in the transition [3].

10. Expect the home robot later, and by way of industry. Playter's route runs through industrial products and mass manufacturing first, with usefulness ahead of charm at every stage [1].

Sources

1. Robert Playter on Boston Dynamics: Atlas, Spot, Stretch and the Business of Legged Robots · Lex Fridman Podcast
2. Dr Matthew Macdougall Neuralink Technologies To Enhance Human Brains · Huberman Lab
3. Kate Darling on Social Robots, the Animal Analogy, Privacy, Manipulation and Lessons from MIT's Epstein Scandal · Lex Fridman Podcast