The Robot Episode: Four Leaders on What's Coming

The transcript records a series of interviews conducted by host “J-Cal” at the Machina conference in Paris, France, described by the host as being about “AI in the real world.” It contains four substantive robotics interviews, plus AppLovin and Plaud sponsor segments.

1. J-Cal — host/interviewer.
2. Dr. “Peter Funkhouser” [transcript spelling/name unclear] — introduced as co-founder and CEO of “Ennebotics” [transcript unclear], maker of “Eni-mo” [transcript unclear].
3. “Bert Bornyck” [transcript spelling/name unclear] — introduced as founder and CEO of “OneX” [transcript spelling unclear], maker of the Neo humanoid robot.
4. Amanda McMaster — introduced as interim CEO of Boston Dynamics.
5. Professor Jonathan Hearst — introduced as co-founder and chief robot/chief robotics officer of Agility Robotics.

The names and company/product spellings above reproduce the transcript rather than independently correcting them.

The host opens from Paris at Machina, describing the event as focused on AI operating in the real world.

An AppLovin advertisement states that:

• AppLovin began with an $8 domain and no venture-capital funding.
• It became one of the world’s largest advertising platforms.
• The same advertising engine now powers AppLovin advertising for e-commerce.
• Advertisements can run inside mobile games and supposedly reach more than one billion people with full-screen attention.
• The platform is described as finding buyers and optimizing campaigns for profit after the advertiser specifies a target.
• A cookware brand is said to have grown from $4 million to $16 million, become profitable, and be “on pace” for $80 million that year.

These are sponsor claims rather than independently established facts.

The host says the guest has worked in robotics for nearly 20 years and that his company has existed for about 10 years. The first five years are characterized as being closer to a research laboratory; the following five years have focused on real deployments.

The company’s product is described primarily as an industrial inspection and data-collection system for critical infrastructure, using a four-legged or “dog” robot as the physical platform.

The guest explains why quadrupeds became a practical form factor:

• Four legs provide strong mobility.
• They can climb stairs and travel through many of the same locations humans can.
• Four legs offer balance and stability.
• A wide footprint provides multiple footholds.
• This matters in difficult industrial environments containing slippery floors, rain, snow, vegetation and other unstable conditions.

The host asks why humanoid robots should not simply use four legs, given concerns about two-legged robots falling. The guest says form factor should follow the application:

• A humanoid may make more sense in narrow human spaces such as a coffee shop.
• Human-height interaction can matter.
• In the industrial facilities his company serves, there is generally sufficient room and the stability of four legs is preferable.

The host jokingly proposes a four-legged/human upper-body design, repeatedly referring to it as a “senator”; the intended mythological term is [transcript unclear].

The host initially says there may be “thousands” deployed, then the number is corrected to hundreds forward deployed over the previous five years.

The robots are described as expensive industrial equipment. The host places their purchase price in the low hundreds of thousands of dollars and speculates that annual service could cost tens of thousands of dollars. The guest agrees with the broad characterization that they are expensive industrial systems, but the service-contract estimate originates with the host.

They are not consumer/home robots. Their value comes from sensors, compute and inspection capability.

The guest explicitly frames the product as being about doing things humans cannot do as well, rather than simply replacing workers.

He says human eyes and ears cannot detect every relevant industrial signal. The robots can carry combinations of:

• thermal cameras,
• acoustic microphones,
• gas-concentration sensors,
• conventional cameras,
• AI processing,
• and other inspection sensors.

Applications include detecting:

• very small gas leaks,
• equipment overheating,
• acoustic abnormalities,
• and other signals humans may not directly perceive.

The guest argues that these machines can therefore perform “superhuman” inspection.

The primary financial justification is avoided downtime. According to the guest, some industrial assets lose hundreds of thousands of dollars per hour when they stop operating. Saving even minutes or hours can therefore pay for an expensive robot.

That economic logic also makes it reasonable to put expensive sensors and powerful GPUs on the machine.

The guest gives a typical mission duration of approximately two hours.

Robots return autonomously to docking stations to recharge rather than depending on people to swap batteries.

A numerical exchange about deployment frequency is noisy. The transcript first records the guest saying some customers perform missions “40 times a day,” followed by an exchange of “14? 14. 40, 4 is 0.” The intended number should be verified.

One use case involves observing a specific moment when an “electric art furnace” [transcript unclear] operates. The customer wants data precisely during that event because it may be too dangerous to place a human nearby.

The guest says customers do not ultimately care about owning an impressive robot. They want data and actionable insight. The physical robot is simply the mechanism for gathering that data consistently.

The guest says processing is split between the robot and the cloud.

Functions that must operate locally include:

• obstacle avoidance,
• real-time navigation,
• ensuring data quality,
• and functions that cannot depend on reliable connectivity.

He gives the example that if a blurry image is uploaded before its poor quality is noticed, it is already too late to recollect that particular observation.

Cloud systems can then perform:

• contextual analysis,
• historical comparisons,
• downtime analysis,
• and broader interpretation.

The guest does not see a universal need for continuous 24-hour missions. The transcript refers to a maximum in an “eight-hour range” with time reserved for charging, although exactly what operational window is being defined is somewhat unclear.

He argues there are diminishing returns to inspecting increasingly often. If a site was previously inspected manually once or twice per day, increasing that to eight, fifteen or twenty inspections can already represent a major improvement.

That increased frequency also keeps people away from hazardous locations.

The guest identifies offshore infrastructure as one of the most interesting deployments.

For offshore oil operations, human access may require helicopter trips costing tens of thousands of dollars per flight, according to the guest.

Offshore installations also have:

• limited personnel,
• difficult accessibility,
• and strong requirements for continuous operation.

The technology is also used around offshore wind infrastructure.

The robots do not walk underwater. Instead, the guest describes robots operating at offshore transformer/converter stations to which many turbines connect. The transcript says these stations transform AC to DC before transmission, which should be independently checked.

The guest says systems operate:

• in Norway at approximately –20°C,
• and in deserts at approximately +40°C, +50°C or +60°C.

He emphasizes temperature, dust and humidity as conditions in which robots can be preferable to people.

Most notably, he says the company has built a robot for explosive atmospheres in industries such as oil, gas and chemicals.

Where methane or similar gases may be present, creating a spark can be dangerous. He describes this robot as being designed so it will not generate such a spark.

This supports his broader principle: machines should be sent into dangerous environments where possible instead of people.

The host raises the possibility of sending similar robots to the ocean floor.

The guest says his company does not currently do that and notes that submarines and different robotic form factors serve such environments.

He nevertheless agrees with the general principle that robots should work in places that are:

• dangerous,
• remote,
• or repetitive/boring.

The discussion notes that other organizations are developing surface and subsurface monitoring robots, including military applications.

The host asks when an inspection robot that discovers a gas leak will also be able to repair it.

The guest says this is not realistically available yet.

Manipulation can be demonstrated, but achieving something like 99.9% reliability, particularly in explosive and harsh environments, remains difficult.

Nearer-term manipulation tasks include:

• closing levers,
• opening cabinets,
• and progressively interacting with industrial machinery.

The longer-term goal could involve more complex manipulation and potentially robots with three or four arms.

The guest contrasts industrial reality with demonstrations of humanoids folding laundry:

• laundry demonstrations happen in highly controlled settings;
• an outdoor hailstorm or freezing industrial location creates far greater perception and manipulation difficulty.

He expects AI and engineering progress eventually to solve these problems but says substantial work remains.

When asked what percentage of the company's robot is sourced from China, the guest answers zero percent.

He says this was largely the result of the company’s historical development rather than a direct ban:

• it sourced locally,
• obtained chips from the United States,
• and built much of its own architecture because the company began a decade earlier.

He nevertheless says modern robotics companies should make use of global hardware commoditization where economically sensible.

The relevant distinction is between:

• critical, expertise-heavy components,
• and simpler materials or assembly.

Asked where important components can be sourced, the guest says China is still the strongest region pushing many components, while good companies also exist in Europe and the United States.

Routine labor assembly can happen elsewhere, but critical component expertise is more important than assembly location.

The host asserts that China has stolen intellectual property at scale, speculates that the guest's IP may already have been stolen, and predicts Chinese robots could be 80% cheaper. These are host claims/speculation rather than guest-confirmed facts.

The guest's answer focuses instead on differentiation.

He says a Chinese quadruped may be impressive hardware capable of walking or performing backflips, but customers need much more than walking hardware.

His company's solution includes:

• autonomy,
• inspection,
• intelligence,
• workflow integration,
• software and services,
• and trusted data handling.

He also emphasizes cybersecurity and says the company has ISO certification related to such requirements.

Critical-infrastructure customers may be uncomfortable placing numerous cameras and sensors inside a facility if they do not trust who ultimately controls the data.

The host jokes about nuclear-plant information going to the Chinese Communist Party. The guest's more general point is that data leakage is a real concern.

The host asks whether Europe may need to build more of its own military technology, especially if it cannot always assume U.S. support.

The guest personally agrees that Europe has a responsibility to build enough technology to remain independent.

However, he says his company is not currently developing military products and does not presently intend to do so.

He stresses that converting an industrial quadruped into a military product is not as simple as using the same chassis.

Military applications require different:

• communications,
• autonomy,
• latency requirements,
• human-in-the-loop systems,
• and application software.

The company receives military-related requests, according to the exchange, but the guest argues merely shipping an existing robot does not solve a meaningful military problem.

A serious military product would require dedicated effort and, in his view, effectively a different team.

Industrial inspection autonomy can be planned and configured ahead of time. Military autonomy may require decisions and communication on a millisecond scale.

Asked about armed quadrupeds, the guest says:

“I personally don't like it. I hate it.”

He argues engineers should take pride in creating technology for good.

He distinguishes between defense and actively mounting weapons on robots, which he considers risky.

He also argues current robotic technology is not mature enough to casually entrust with actions that could put people in danger.

Regarding videos of weaponized robots, he believes many are still early demonstrations rather than mature deployed systems.

He contrasts this with drones in Ukraine, which he describes as a much more mature technology category that entered warfare rapidly out of necessity.

The guest says that about four years earlier, his company, Boston Dynamics and other robotics companies signed or wrote a letter condemning robot weaponization.

The reason, he says, was that engineers did not want their work used that way and considered weaponized robots dangerous.

The host asks whether Neo will actually ship in 2026.

The guest says it will.

He immediately qualifies that launch volume will be small:

• only a handful of customers are expected to receive Neo in 2026;
• deployment will start slowly;
• the company wants to manage expectations and “do it right.”

The company originally offered two payment structures:

• an early-adopter full upfront purchase,
• and a subscription.

The guest says the subscription model is still evolving and will partly be shaped alongside early customers.

The guest says the company plans, within the coming weeks, to open Neo so other developers can build on it.

He describes something resembling:

• an app store,
• or “skill store,”

in which developers could create capabilities for the robot.

The host uses a salad-making skill as an example: a developer could create a capability allowing a customer’s Neo to make a salad, and the customer might purchase or subscribe to that capability.

The guest confirms such consumer skills are part of the concept.

However, he stresses that the company’s ambition extends well beyond household robots.

The larger mission is to create an “abundance of labor across society” using humanoids.

An open developer ecosystem could therefore:

• extend Neo into commercial and industrial environments,
• create more consumer capabilities,
• improve the platform,
• and accelerate progress toward autonomous household agents.

The guest says the company has not released an official total preorder number.

He does state that the first 10,000 units sold out in the first few days.

When the host asks about subscription pricing, $500 per month is confirmed in the exchange.

Both speakers characterize the earliest customers as unusually early adopters rather than ordinary mass-market buyers.

The guest explicitly warns that owning a home humanoid in 2026 will be “rough around the edges.”

The host says the robots will fall; that statement should not be treated as a precise product-performance claim from the guest.

The guest says performance has advanced faster than the company expected.

Although he stops short of guaranteeing it, he believes Neo may ship with an experience very close to full autonomy.

He says the company deliberately did not promise that initially because it was too early.

His current assessment is that a useful, fully autonomous experience may be possible for customers who accept limitations.

For users who expect everything to work immediately and reliably on day one, some:

• teleoperation,
• or system guidance

may still be necessary.

He says the trajectory toward useful full autonomy is now clearer than it was earlier.

The host gives an anecdote about a New York chicken-sandwich business allegedly using a cashier in Manila through Zoom because it could not find a local cashier.

The host states she earned approximately $3 per hour, describing that as a high cashier salary in the Philippines. This anecdote and wage characterization are host claims requiring verification.

He uses the example to imagine remote workers taking control of humanoid robots to perform physical tasks.

The guest says he expects all of these operating models to exist.

The guest describes a possible commercial implementation.

An operator could purchase a fleet of Neo robots and receive:

• the robots,
• the robot operating system,
• fleet-management tools,
• and data-collection equipment.

That data equipment includes gloves using the same tactile sensors as Neo and a corresponding vision system.

A business could then:

1. gather task data in its own environment;
2. feed the data into the company's system;
3. have the data densely captioned/processed;
4. fine-tune a model;
5. deploy that model back into the robots;
6. operate an automated business workflow.

If autonomy does not handle every case, a human can intervene through teleoperation.

Those interventions also produce additional data, potentially improving the model.

At the opposite extreme, a business could choose to use the robot almost entirely through teleoperation.

The guest therefore sees a spectrum rather than one fixed autonomy model.

The platform will also allow outside foundation models to run on Neo.

The host mentions OpenAI and Claude as examples of possible outside model providers; those specific integrations are hypothetical examples from the host, not announced partnerships.

The guest says “100%” when asked whether external models will be allowed.

At the same time, he says he strongly believes the company's own model will be best.

His reasoning is that a company controlling the stack from manufacturing through the finished product should be able to produce a highly competitive model; if it cannot, he considers that a failure.

Nevertheless, he says he believes in competition.

The guest says there is currently no single general robotics model that solves every task.

He does not want his core research team spending the next several years implementing individual customer ERP integrations and similar site-specific projects.

Instead, the company wants to focus on the general problem:

How can embodied AI become capable enough to create an abundance of labor?

Opening the platform allows other developers to handle specific applications while the core company works on generalized intelligence.

He argues that a large robotics ecosystem benefits all participants.

The host speculates that relatively few human operators might supervise many robots and only intervene occasionally. A numerical passage about how many robots each team could monitor is [transcript unclear] and should not be published as a clean statistic.

The guest then describes a personal teleoperation use case.

He spends time in Norway and the San Francisco area and travels frequently. He would like to use Neo as a physical avatar inside his company while away.

He describes being able to:

• move around,
• inspect physical parts,
• speak with colleagues,
• and participate in meetings

through the robot.

He says this type of telepresence is valuable enough that he believes it will remain useful even if autonomy becomes extremely capable.

The host characterizes it as having an avatar at a remote factory.

Another enduring teleoperation use case is remote infrastructure.

A power station could have a robot waiting at a location an hour or more from the nearest technician.

If a rare fault happens only every few months, fully automating that exact procedure may not be economically sensible.

Instead, a remote expert could immediately take control of the robot, inspect the location and operate switches or other equipment.

The guest connects this with the concept of “expert in place”: effectively teleporting an expert anywhere in the world.

The host gives a surgeon as another conceptual example.

The guest says Neo's newer hands have become capable enough that conventional teleoperation can no longer exploit all of their hardware capability.

A remote operator does not naturally experience all the tactile information available to the robot.

Solving this with complete haptic-feedback systems creates additional complexity and can make the interface slow or cumbersome.

Therefore, for learning basic dexterity, the company increasingly prefers humans to perform tasks while wearing robot-equivalent sensors in as unobtrusive a way as possible.

That produces natural human behavior data without requiring a human to manually puppeteer the robot.

The guest describes a long-term company bet:

If the physical robot is sufficiently similar to a human, it can learn from the enormous existing body of video showing humans interacting with the world.

He says the company is beginning to see promising evidence that this approach works and connects that progress to its World Model Lab.

A phrase about having the “scaling loss” [transcript unclear] should be checked before publication.

The guest describes a hierarchy of robotics training data.

• Smallest amount.
• Highest task-specific quality.
• Operators attempt tasks quickly and accurately.
• Failed attempts may simply be repeated.
• The company selects strong examples in which execution approaches human quality.
• Relatively little of this data is needed because its purpose is primarily model alignment/fine-tuning.

Humans wear sensors corresponding closely to the robot's sensing system.

This dataset can be larger.

It is not physically produced by the robot itself but remains close to the robot's embodiment.

This is video from a person's viewpoint.

It is further removed from direct robot data but still useful because Neo's hands and physical arrangement are designed to resemble human hands and motion.

This is the enormous base layer.

It includes YouTube and general video of humans and the physical world.

The guest describes this dataset as “absolutely ludicrously immense” relative to datasets that robotics companies can collect directly.

The guest argues that achieving broad physical intelligence will require multiple orders of magnitude more data than companies can realistically collect in the next couple of years solely through egocentric video or specialized sensor capture.

He offers a general AI principle:

Major model breakthroughs have repeatedly occurred when researchers found a way to unlock a huge new source of previously unusable data.

For robotics, he believes that source must be the immense quantity of existing general human video.

This creates a physical-design requirement: Neo needs to be close enough to human embodiment that human video transfers meaningfully.

The guest says seemingly small mechanical details matter.

Human skin, flesh and tissue behave nonlinearly. A robot intended to transfer learning from human video needs to consider factors such as:

• deformation under force,
• friction,
• impact energy when contacting surfaces,
• hand proportions,
• arm length,
• and other physical geometry.

The host introduces an NBA claim that players with unusually long wingspans may earn 20% more for a few extra inches of reach. That figure is the host's assertion and requires verification.

The guest says the company jokingly refers to arm-length proportions as the “gorilla coefficient.”

The guest says a major differentiator for his company is that its main cross-embodiment reference is the human, rather than another robot.

The company is therefore investing heavily in pretraining its models on internet video involving humans.

He acknowledges that direct robotics data will eventually be superior because it contains:

• actions,
• tactile signals,
• force data,
• and other robot-specific information.

The catch is that large amounts of robot data can only be generated after many useful robots have been deployed.

Therefore the company first needs a base model capable enough that customers will pay to deploy robots throughout society.

Those deployed robots can then gather the much richer robotics data needed for subsequent generations.

The host asks when robots will reach a point where they recursively improve, build other robots and pursue goals rather than being manually instructed task by task.

The guest says the concept is not science fiction and that the company thinks seriously about it.

He also warns that forecasting is difficult because progress has repeatedly been faster than he expected.

A year earlier, he says, he would have been more pessimistic about where AI would be at the current point.

He nevertheless makes a very strong prediction:

• he is “extremely sure” hard takeoff is less than a decade away;
• his current personal bet is three years.

He defines hard takeoff in physical terms:

• robots building robots,
• robots building data centers,
• robots building chip fabrication facilities,
• robots performing mining,
• robots performing refining,
• and the emergence of a self-sufficient physical production system creating a true abundance of labor.

He considers the difference between three years and ten years relatively small in the longer history of humanity.

The guest argues that purely digital intelligence cannot create its own physical substrate.

To become self-sufficient, AI must be capable of operating in the physical world.

He sees scientific research as an important example.

The company is reportedly receiving demand from organizations wanting to automate laboratory work.

His reasoning is:

• an AI system can propose experiments digitally;
• but if it cannot physically build experiments, execute them and observe the results, its ability to autonomously advance science remains constrained.

He predicts that as AI becomes physical, these systems will increasingly perform scientific work.

Although he will not commit to an exact date, he says the relevant horizon is “years, not decades.”

The interview closes with the guest saying customers can order Neo directly through the company's website.

The host compares that simplicity and transparent pricing to ordering a Tesla online.

A Plaud advertisement argues that exact wording in business conversations can matter, citing examples such as numbers stated during diligence calls or commitments made during board meetings.

The advertisement says Plaud:

• captures conversations,
• turns them into searchable intelligence,
• allows users to ask questions of recorded material,
• and reduces dependence on manually scrolling through recordings or memory.

The phrase “get the answer with no receipt” appears in the transcript and is likely transcription noise; the intended wording is [transcript unclear].

The host recounts Boston Dynamics' ownership history as:

• originally independent,
• acquired by Google under Sergey Brin and Larry Page,
• later owned by Masayoshi Son/SoftBank,
• and now owned by Hyundai.

Amanda McMaster confirms that the host has the broad history correct.

She says Boston Dynamics is no longer primarily an AI or R&D laboratory.

Its current focus is deploying robots into real-world customer environments.

The first major commercial platform discussed is Spot, Boston Dynamics' four-legged industrial robot.

McMaster says Spot is performing real work at customer sites and states that Boston Dynamics has:

• more than 500 customers
• in more than 46 countries.

She further claims Spot is the most deployed and most utilized mobile autonomous robot on the planet.

That superlative should be independently verified before publication.

Industrial inspection is described as the leading application.

Spot can perform:

• acoustic inspection,
• gauge reading,
• vibration detection,
• and monitoring of expensive industrial assets.

The same robot can perform inspection during the day and security/perimeter work at night.

The commercial turning point, according to McMaster, was demonstrating customer ROI rather than merely showing technically impressive robots.

Boston Dynamics wants industrial customers to see meaningful return on investment in under two years.

The company now emphasizes useful work rather than entertainment demonstrations such as dancing robots.

The host notes that inspections previously depended on people when they were performed at all.

He emphasizes that humans can make mistakes.

Spot can repeatedly traverse facilities such as:

• water-treatment plants,
• bridges,
• pipelines,
• and other infrastructure,

while collecting multiple sensor streams.

The exact full sensor list given by the host includes some assumptions, but acoustic and vibration-related inspection are explicitly discussed by McMaster.

Boston Dynamics initially chose a CAPEX purchase model for Spot.

McMaster says this works well because industrial customers already think of equipment and tools as capital purchases.

She says price depends on configuration:

• approximately $100,000 for a base Spot,
• up to approximately $300,000 for a fully configured deployment including services and integration.

The host characterizes that as a range from a Tesla to a Ferrari.

For Atlas, McMaster says Boston Dynamics will likely use a robot-as-a-service model, since customers may want to expand and contract humanoid capacity over time.

The host suggests an industrial robot could have a useful life exceeding five years. This is primarily his framing rather than a specific durability guarantee from McMaster.

McMaster says Boston Dynamics measures mean time between intervention and has exceeded approximately 3,000 hours.

Spot has a charging station.

Its battery operates for roughly 90 minutes.

The transcript suggests sites may use two units so one robot can work while another charges, though exact operational arrangements may depend on customer configuration.

Spot docks itself rather than automatically replacing its own battery.

McMaster says Atlas has swappable batteries and corrects the host when he assumes a human must perform the swap.

She says Atlas can perform the operation itself.

The transcript describes Atlas as having two batteries and turning its torso while replacing one battery while retaining another as backup. The precise mechanism should be verified from technical material because the wording is somewhat compressed.

The host estimates that nearly continuous operation could reduce effective robot operating cost to only a few dollars per hour.

McMaster says Boston Dynamics has not framed Spot principally as labor replacement.

Two reasons are given:

1. some inspections were not actually being performed by workers consistently, even if they were nominally assigned;
2. the company would rather shift human employees toward higher-value knowledge work than send them repeatedly into inspection environments.

She acknowledges labor replacement can be one ROI metric.

However, Boston Dynamics increasingly asks a broader question: How much financial loss did the robot prevent?

An example is given of discovering an air leak that otherwise could have created approximately $3 million per day in losses. The origin/customer specifics of this figure are not supplied and should be verified.

McMaster acknowledges that labor displacement is part of the market discussion but says it should not be the only way robotics value is measured.

Both speakers emphasize removing people from the classic “dull, dirty, dangerous” jobs that may be physically damaging.

McMaster describes the architecture using a simplified two-brain model.

The systems responsible for:

• movement,
• physical reliability,
• manipulation,
• and dynamic control

run directly on the robot.

She identifies this physical intelligence as a traditional Boston Dynamics strength.

The higher-level reasoning system responsible for semantic understanding of the environment can operate in the cloud.

She says Boston Dynamics may:

• build some of this itself,
• and partner with outside AI providers.

The transcript names “Google Devine” [transcript unclear] as a possible partner/reference and should not be published without checking the intended organization/product.

A third layer contains customer-specific knowledge such as:

• workflows,
• processes,
• tools,
• and how the robot should interact with a particular facility.

That layer could live:

• on the robot,
• in the cloud,
• or elsewhere in the computing architecture depending on need.

The host asks what percentage of Boston Dynamics' robot is built either in the United States or outside China and Taiwan.

McMaster replies “100% of the robot.”

Because the wording of the question combines U.S. production with broader non-China/Taiwan sourcing, this answer should not be simplified into “100% made in America” without further verification.

Asked whether Chinese humanoid robots should be allowed into the United States, McMaster answers no.

Her reasoning includes:

• safety,
• reports of information from some quadruped robots being sent or “back-channeled” to China,
• protection of U.S. intellectual property,
• avoiding strategic dependence,
• and building robotics manufacturing in the United States and allied countries.

She compares robotics strategically to semiconductors and argues the United States should establish a coordinated national robotics strategy.

Boston Dynamics, she says, participates in some governmental discussions around the issue.

She wants more U.S. robotics companies to join that effort.

The host extends this into a geopolitical argument that the United States must not only build competitive robots but persuade the rest of the world to use an American rather than Chinese robotics ecosystem.

That stronger “existential” framing belongs primarily to the host.

McMaster says Boston Dynamics has publicly maintained an anti-weaponization position.

Asked why, she frames the current rationale primarily around business focus:

• industrial applications are the company's priority;
• military weaponization would distract resources;
• Boston Dynamics needs to focus on markets it expects to win.

When the host asks whether this is philosophical or strategic, she suggests the answer may differ depending on whom one asks, but speaking in her leadership capacity she emphasizes commercial focus.

Boston Dynamics does work with government customers on non-weaponized applications.

McMaster specifically confirms current work involving EOD — explosive ordnance disposal.

The discussion also presents hypothetical non-weapon roles such as:

• retrieving a wounded soldier,
• carrying medical supplies,
• and bomb disposal.

McMaster clearly confirms EOD as an actual current application; the other examples arise in the host's framing.

The host presses McMaster on whether Boston Dynamics would build weaponized systems if geopolitical competitors deployed armed quadrupeds.

McMaster says this is a difficult question that the company may have to answer if that situation arises, though she hopes it never does.

She does not commit to building armed robots.

The host then speculates that a future president could effectively compel Boston Dynamics to do so and jokes that refusal might lead to nationalization. This is rhetorical speculation from the host.

The host also says Chinese systems have been demonstrated with AK-47-style weapons.

McMaster makes clear those demonstrations do not involve Boston Dynamics robots.

She describes the prospect of weaponized robots as “terrifying.”

She nevertheless says that if circumstances eventually forced Boston Dynamics to make a difficult decision, the company would make what it believed was the right decision. For now, she intends to keep the company focused on commercially attractive applications.

Near the end, the host makes a striking unsupported assertion:

“The CIA, the FBI, and the Department of War have many of your robots with many weapons attached to them currently.”

The host immediately frames it jokingly with “Don't tell anybody.”

McMaster does not confirm this claim in the transcript.

This must be treated as an unverified host allegation/joke, not a factual statement.

The reference to a U.S. “Department of War” should also be checked carefully against the date and intended agency name before any publication.

McMaster says Boston Dynamics is based in Waltham, outside Boston.

The company permits some remote work.

She also says Boston Dynamics is considering establishing a presence on the U.S. West Coast, in light of robotics/AI talent concentrated around the Bay Area.

The host introduces Hearst as having a robotics PhD from 2008 and more than two decades in the field. The arithmetic/timeline should be checked because the exact “over 20 years” wording may include pre-PhD work.

Hearst says that around 20 years earlier:

• industrial robotics was largely conventional automation,
• while humanoids and autonomous mobile robots remained predominantly research subjects.

Researchers were working on both:

• physical capability,
• and the intelligence required for autonomous machines.

He says that work is now breaking into the real world and producing direct practical impact.

Universities have responded to robotics growth and student interest by expanding programs substantially.

Hearst draws a distinction between looking human and being useful in human spaces.

He says it has been comparatively easy for a long time to build something that resembles a person.

The hard problem has always been making that robot perform useful work in real human environments.

He believes robotics is now beginning to cross that threshold.

A robot does not need perfect human appearance. What matters is whether its humanoid characteristics allow it to perform useful work.

The host argues that large language/multimodal models have effectively made certain forms of perception available much more cheaply.

A robot can increasingly recognize everyday objects rather than needing each category narrowly programmed.

Hearst agrees that perception was historically extremely difficult and says its dramatic improvement is a major inflection point.

He also says AI has allowed people to see a plausible path toward generality: robots becoming context-aware enough to perform many different useful activities.

However, he rejects the idea that all robotics intelligence is therefore “free.”

The host uses his CafeX investment as an example of older automation:

• a robot arm can reliably make coffee or pour beer,
• but the operation historically had to be manually programmed.

He asks whether large models trained on internet videos and recipes now solve the control problem.

Hearst answers:

“Not yet.”

Language models have something approaching an internet-scale training corpus.

Equivalent data does not exist for physical robot control.

There is no enormous internet dataset containing, for every real-world situation:

• the appropriate torque command for every motor,
• paired with all sensor observations,
• and the correct physical action.

That control data has to be created.

Hearst says it is misleading to treat AI as one monolithic system.

Instead, robotics uses many different computational tools.

Methods for generating robot-control capability include:

• learning from demonstrations,
• teleoperation,
• animation,
• motion-capture data,
• and other forms of training input.

Each contributes something different.

Human teleoperation has its own ceiling.

A human controlling a robot may not discover the mechanically optimal way for that robot to move.

Robot bodies may have capabilities and dynamics different from humans.

Therefore the robot also needs opportunities to practice.

This motivates:

• world models,
• simulation,
• sim-to-real transfer,
• and physical-world learning.

The host describes world models as simulated environments in which a robot could practice enormous numbers of tasks without breaking physical objects.

Hearst agrees world models are part of the solution but rejects any idea that they are a silver bullet.

Simulation never perfectly reproduces reality.

Sources of error mentioned include:

• fluid or “wave dynamics” [transcript wording should be checked],
• condensation on a glass,
• incomplete models of robot dynamics,
• and countless physical irregularities.

Therefore significant learning still has to happen using real robots in real environments.

The host asks whether robotics will reach a crossover point where recursive practice rapidly creates general capability, analogous to self-play systems learning games such as chess or Go.

Hearst says he does not believe in one singular moment or one technological silver bullet.

His preferred analogy is a snowball gathering momentum downhill.

Progress accelerates because increasing amounts of:

• capital,
• engineering effort,
• research,
• and experimentation

are being投入 into many approaches simultaneously.

Hearst emphasizes that humans evolved specifically to learn effectively from very small amounts of data.

A person can often:

1. watch a task briefly;
2. understand it;
3. practice;
4. and improve quickly.

Robots currently require vastly more data and examples.

The field is still learning how to teach robots to learn efficiently.

Robots eventually have a powerful advantage humans do not.

If a person learns violin, that skill cannot simply be uploaded directly into another person's nervous system.

With robots, a skill learned by one unit can potentially be distributed digitally to:

• every robot of the same type,
• and then adapted for related robot hardware.

Hearst says that once robot learning improves, networked transfer could therefore make each machine's experience much more scalable.

The host describes the currently deployed generation as approximately Digit 4.0 and says Digit 5.0 is coming. Those version labels should be verified.

Hearst describes current work as multipurpose but still relatively scoped.

Examples include:

• picking up bins and totes,
• carrying them,
• moving goods through warehouses,
• individual-item picking,
• filling bins,
• palletizing,
• and depalletizing.

These tasks help justify the humanoid body:

• two arms are useful for large objects;
• whole-body control allows manipulation;
• balance is necessary when lifting;
• humanlike proportions allow work around tall shelves and narrow human-designed spaces.

Ultimately, however, Hearst argues the real value of humanoids is versatility rather than any one task.

Hearst says Digit V5, expected later that year according to the interview, represents a major scaling milestone.

He describes it as the first time their balancing humanoid will be able to leave a traditional work cell without requiring a physical barrier between the robot and nearby humans while meeting the warehouse safety requirements relevant to its use.

That matters because conventional industrial robots often operate behind safety barriers.

Hearst describes an experience working with Amazon.

The robots successfully performed the desired task and met the R&D objectives.

Agility therefore expected deployment to follow.

Instead, Amazon indicated the system could not deploy because it did not yet meet required industrial safety standards.

Hearst says solving this proved “super hard.”

Over roughly the following two or three years, safety required a bottom-to-top redesign affecting essentially every major system in the robot.

The goal was to create a machine safe enough to work in proximity to people rather than inside isolated robot zones.

Hearst declines to discuss Agility's current bill of materials.

He does say manufacturing costs are falling and predicts robots will eventually sell in the vicinity of automobile prices.

He argues that purchase price is less important than lifetime economic output.

A machine working most of the day for several years can generate substantial total value.

Hearst confirms a target of approximately 20 working hours out of every 24 for Digit V5, enabled by improvements in fast charging.

The host calculates:

• roughly 7,000–8,000 productive hours per year,
• around 8,000 hours as an illustrative round number,
• multiplied over five years,
• giving approximately 40,000 working hours.

That 40,000-hour figure is the host's illustrative arithmetic, not necessarily a warranted service-life specification.

The host suggests humanoid robots could ultimately cost $20,000–$40,000.

Hearst says prices may eventually reach that sort of level, but meaningful cost reductions require manufacturing scale.

He specifically says the industry needs something like 100,000 robots deployed before those economics become realistic.

The host then turns a hypothetical $40,000 capital cost divided across 40,000 hours into roughly $1 of capital cost per working hour.

That calculation excludes many other operating costs and should not be published as actual robot hourly cost.

The host compares future robot economics with human factory wages of roughly:

• $20,
• $30,
• or $40 per hour

in Western economies, while mentioning lower wages elsewhere.

Hearst agrees this leaves a large commercial market but does not endorse each wage figure as a universal benchmark.

Agility supports both:

• direct CAPEX sales,
• and robot-as-a-service arrangements.

Robot-as-a-service lowers the customer's initial commitment and risk.

Hearst declines to disclose Agility's current hourly service price.

Hearst makes an important economic argument:

• robot capability is improving, increasing the amount of value each machine can produce;
• manufacturing cost is falling at the same time.

The economic value of labor robots is ultimately linked to what customers would otherwise have to pay people to perform the same work.

He therefore describes the market price/value relationship as relatively inelastic for a long period, because robots can become cheaper to make while still replacing or augmenting much more expensive labor.

Asked when more than half of factory workers will be robots, Hearst responds that in many applications the majority of work is already performed by automation.

He points to:

• autonomous mobile robots,
• conveyor systems,
• industrial robotic arms,
• and other machines.

Humanoids are therefore another form of automation rather than the beginning of automation.

In the United States, he argues economic growth cannot depend solely on population growth.

Productivity and GDP growth require increasing efficiency and automation.

The host adds a comment about anti-immigration politics reducing labor-force growth; that political interpretation is the host's, not Hearst's core argument.

Hearst expects automation to continue until some package-processing systems become effectively lights-out:

• an autonomous truck arrives;
• packages are automatically sorted;
• another autonomous truck departs;
• almost no human intervention is required.

However, he says such a tightly optimized environment may use specialized automation rather than humanoids.

If one task happens continuously and predictably 24/7, a dedicated machine can be more efficient.

Humanoids are most valuable where machines need to enter environments designed for people and perform human workflows.

As older factories become automated, large numbers of other human environments still remain.

Agility is working or exploring applications in:

• retail,
• grocery stores,
• hospitals,
• construction sites,
• and last-mile package delivery.

These environments remain designed around human movement and therefore potentially favor humanoid form factors.

The host imagines an autonomous vehicle arriving and a humanoid robot stepping out to carry a package to a customer's door.

Hearst says Agility explored essentially that use case with Ford very early in Digit's development.

He says a video exists showing an early Digit robot:

• getting out of a vehicle,
• walking to a front porch,
• navigating stairs,
• and delivering a package.

He estimates this demonstration occurred approximately seven years earlier.

The capability remains on the roadmap.

However, he does not consider it the best initial market because easier and larger near-term industrial opportunities already exist.

Asked for overlooked positive applications, Hearst says simply moving physical objects is itself an enormous opportunity.

It can remove people from jobs embodying the traditional robotics 3Ds:

• dull
• dirty
• dangerous

He hopes future generations will look back on some current manual jobs the way modern society looks back on extremely dangerous historical labor such as early-20th-century coal mining.

His broader claim is that technology can improve quality of life while generating new categories of work that earlier generations could not have imagined.

Hearst continues to teach robotics and confirms the relevant graduate program has more than 100 students.

He sees the current transition as a major opportunity for young engineers.

During periods of technological change, new graduates may have an advantage because professionals trained over the previous 20 or 30 years must also learn the new methods.

He cannot predict exactly what robotics careers will look like ten years from now.

He recommends developing strong core engineering capabilities because those skills should remain transferable.

The host asks about trade-level or technician careers comparable with:

• electricians,
• HVAC technicians,
• carpenters,
• or contractors.

Hearst says those opportunities will exist.

Examples include:

• robot operators,
• robot assembly,
• manufacturing,
• deployment,
• field operations,
• and maintenance.

He notes that robots cannot yet manufacture and maintain the entire robotics ecosystem themselves.

The speakers joke about calling robot maintenance “clanker maintenance.”

A suggestion that “clanker” is a Disney trademark is speculative/joking and should not be treated as established fact.

Using General Grievous from Star Wars as an analogy, the host asks why a useful robot should not have four or six arms.

Hearst answers from first principles.

The goal is to build the simplest machine that can perform the required work.

• One arm restricts the machine to relatively small items.
• Two arms allow it to lift larger objects.
• A third or fourth arm adds mechanical and control complexity.

Unless additional arms enable a valuable task that two arms cannot perform, the added cost and coordination burden are difficult to justify.

He does not rule out additional arms in future systems but says the design must be driven by a genuine use case.

Hearst names Wall-E as a favorite fictional robot and also discusses EVE and Baymax.

He likes the idea of robots continuing to help and perform the work they were designed to do.

The Baymax discussion turns into a safety analogy: when the fictional robot's programming/memory changes, its behavior becomes dangerous.

Hearst connects that to real robotics safeguards.

Industrial robots require systems such as:

• supervisory circuits,
• emergency stops,
• and formal safety engineering.

The objective is to build machines that cannot injure people even when something goes wrong.

Hearst says Agility Robotics has operations in:

• Salem, Oregon, where the company began and where he is based;
• Fremont, California, where a new facility is opening and substantial robot-behavior development will occur;
• Pittsburgh, Pennsylvania, near Carnegie Mellon.

The host closes by encouraging people concerned about future employment to consider advanced robotics education.

Speaker: Peter [surname transcript unclear]
Exact quote: “Actually, they don't even want the robot. They want the data. They want the insights. The robot is a means for an end to collect the data precisely.”
Context: Explaining why industrial customers deploy inspection quadrupeds.
Topic: Industrial robotics business value.

Speaker: Peter [surname transcript unclear]
Exact quote: “For us it's not about labor replacement, right? It's what can we do better? What can we do superhuman?”
Context: Describing sensors that can detect signals humans cannot directly perceive.
Topic: Robotics versus human labor.

Speaker: Peter [surname transcript unclear]
Exact quote: “Robots should work in environments where people shouldn't be. Dangerous, remote, right?”
Context: Discussion of offshore and underwater applications.
Topic: Hazardous-environment robotics.

Speaker: Peter [surname transcript unclear]
Exact quote: “You know, in a demo, yes. But in reality, getting it to 99.9% reliability in explosive atmosphere... that's still in development.”
Context: Asked when inspection robots will physically repair equipment.
Topic: Reliability gap between demonstrations and deployment.

Speaker: Peter [surname transcript unclear]
Exact quote: “I personally don't like it. I hate it.”
Context: Asked for his reaction to weaponized quadrupeds.
Topic: Robot weaponization.

Speaker: Peter [surname transcript unclear]
Exact quote: “As an engineer, you should have pride, right, to build technology for good.”
Context: Explaining opposition to offensive weaponization.
Topic: Engineering ethics.

Speaker: Bert [surname transcript unclear]
Exact quote: “There will be a handful of customers that get their Neo in 2026.”
Context: Managing expectations around initial shipments.
Topic: Humanoid commercialization.

Speaker: Bert [surname transcript unclear]
Exact quote: “OneX has always been about how do we create an abundance of labor across society through these humanoids.”
Context: Explaining why Neo is broader than a consumer home robot.
Topic: Humanoid economic impact.

Speaker: Bert [surname transcript unclear]
Exact quote: “Getting a home humanoid in 2026 is going to be rough around the edges.”
Context: Warning early buyers not to expect a mature appliance.
Topic: Consumer humanoid readiness.

Speaker: Bert [surname transcript unclear]
Exact quote: “Will we allow other people to build on this 100%.”
Context: Confirming that third-party AI and developers will be able to use Neo.
Topic: Open robotics platforms.

Speaker: Bert [surname transcript unclear]
Exact quote: “Currently, if you look at where the field is, there is no one general model that solves everything for robotics. It's not there yet.”
Context: Explaining why 1X wants a broader developer ecosystem.
Topic: Embodied AI limitations.

Speaker: Bert [surname transcript unclear]
Exact quote: “That's one application of teleoperation that I think actually will never go away.”
Context: Describing using Neo as his physical avatar while traveling.
Topic: Teleoperation.

Speaker: Bert [surname transcript unclear]
Exact quote: “The model is going to be as good as the data.”
Context: Introducing the company's robotics-data pyramid.
Topic: AI training.

Speaker: Bert [surname transcript unclear]
Exact quote: “The bottom layer in the pyramid, which is this video data, general video data, is absolutely ludicrously immense compared to anything else.”
Context: Comparing internet-scale video with specialized robot datasets.
Topic: World models and training data.

Speaker: Bert [surname transcript unclear]
Exact quote: “Ultimately what differentiates one X from all of the other robotics companies is that we are all in on pre-training our own models on this video data on the internet.”
Context: Explaining the company's human-video strategy.
Topic: 1X technical strategy.

Speaker: Bert [surname transcript unclear]
Exact quote: “Our cross-embodiment is not another robot, our cross-enbordial. embodiment is the human.”
Context: Explaining why Neo is designed to resemble human physical behavior.
Topic: Human-to-robot learning.
Note: Quote preserves transcript wording; “cross-enbordial” is [transcript unclear].

Speaker: Bert [surname transcript unclear]
Exact quote: “I am extremely sure there were less than a decade away from hard takeoff.”
Context: Forecasting self-sufficient robotics and automated physical production.
Topic: AI/robotics timelines.
Note: Grammar reflects transcript.

Speaker: Bert [surname transcript unclear]
Exact quote: “My current bet would be three years.”
Context: His present estimate for hard takeoff after giving under ten years as a broader bound.
Topic: Robotics forecasting.

Speaker: Bert [surname transcript unclear]
Exact quote: “All of these things will happen in the coming years as AI becomes physical and exact timeline is a bit hard, but it's years, not decades.”
Context: Discussing autonomous labs and physical scientific experimentation.
Topic: Embodied AI.

Speaker: Amanda McMaster
Exact quote: “It's not a research and development company anymore. We're now focused on real world deployment.”
Context: Explaining the commercial shift at Boston Dynamics.
Topic: Robotics commercialization.

Speaker: Amanda McMaster
Exact quote: “At this point, we have over 500 customers, over 46 countries.”
Context: Describing Spot's commercial footprint.
Topic: Boston Dynamics deployment.

Speaker: Amanda McMaster
Exact quote: “The real inflection point was customer ROI.”
Context: Explaining the transition from impressive demonstrations to valuable industrial work.
Topic: Robotics economics.

Speaker: Amanda McMaster
Exact quote: “We think about two brains, right?”
Context: Introducing the distinction between onboard physical control and higher-level reasoning.
Topic: Robot computing architecture.

Speaker: Amanda McMaster
Exact quote: “No.”
Context: Her direct answer when asked whether Chinese humanoid robotics should be allowed into the United States.
Topic: Robotics sovereignty and China.

Speaker: Amanda McMaster
Exact quote: “We need to have a consorted effort to protect our IP, to make sure that... we are bringing manufacturing of this ecosystem into the United States, we're into our allied countries.”
Context: Explaining her strategic concern about dependence on China.
Topic: Industrial policy.
Note: “Consorted” reproduces the transcript and may be a transcription error.

Speaker: Amanda McMaster
Exact quote: “For what we're trying to do right now in industrial use cases, it's a distraction for our business.”
Context: Explaining Boston Dynamics' present anti-weaponization stance from a leadership perspective.
Topic: Military robotics.

Speaker: Amanda McMaster
Exact quote: “I think that's a tough question. And I think we're going to have to answer it when the time comes and hopefully it never comes.”
Context: Asked whether Boston Dynamics might arm robots if adversaries deploy weaponized systems.
Topic: Military robotics.

Speaker: Amanda McMaster
Exact quote: “It's terrifying. Terrifying.”
Context: Reacting to discussion of armed quadrupeds.
Topic: Autonomous weapons.

Speaker: Jonathan Hearst
Exact quote: “It's very hard to make a robot that can do useful things in human spaces. And we're starting to see that today. And that's the difference.”
Context: Explaining why the current humanoid wave is different from past demonstrations.
Topic: Humanoid robotics.

Speaker: Jonathan Hearst
Exact quote: “That data does not exist for robot control.”
Context: Contrasting internet-scale language-model training with physical-control datasets.
Topic: Robotics AI bottlenecks.

Speaker: Jonathan Hearst
Exact quote: “As with all of these things, there is no silver bullet.”
Context: Discussing whether world models solve robot learning.
Topic: World models.

Speaker: Jonathan Hearst
Exact quote: “I don't believe that there's this singularity. I do believe that things are gonna get better and better. Think of it more like a snowball picking up steam going down a hill.”
Context: Rejecting a single sudden crossover point in robot intelligence.
Topic: Robotics development timelines.

Speaker: Jonathan Hearst
Exact quote: “Robots are not very good at learning yet. Robots take so much more data and so many more examples at a person.”
Context: Comparing human and machine learning efficiency.
Topic: Robot learning.

Speaker: Jonathan Hearst
Exact quote: “One robot learns to play violin, all robots know how to play violin.”
Context: Explaining the long-term advantage of digitally sharing learned skills.
Topic: Fleet learning.

Speaker: Jonathan Hearst
Exact quote: “20 out of 24 hours for our Digit V5 robot.”
Context: Discussing charging improvements and utilization.
Topic: Humanoid economics.

Speaker: Jonathan Hearst
Exact quote: “A humanoid is useful for walking into human environments doing human workflows.”
Context: Explaining why fully automated factories may still use specialized machines rather than humanoids.
Topic: Humanoid form factor.

Speaker: Jonathan Hearst
Exact quote: “The 3Ds of robotics.”
Context: Referring to dull, dirty and dangerous jobs.
Topic: Labor and robotics.

Speaker: Jonathan Hearst
Exact quote: “Anytime there's a time of change like this, students coming out have an advantage.”
Context: Career advice for young people entering robotics.
Topic: Robotics careers.

Speaker: Jonathan Hearst
Exact quote: “It's gonna have to be driven by a real need.”
Context: Explaining why additional robot arms are unjustified without a task requiring them.
Topic: Robot hardware design.