# Robot form factor

> Source: https://aiwiki.ai/wiki/robot_form_factor
> Updated: 2026-07-22
> Categories: Embodied AI, Humanoid Robots, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**Robot form factor** is the physical shape and mobility architecture of a robot: whether it moves on wheels, tracks, legs, or a hybrid of these; how many limbs and joints it carries; and where its manipulators are mounted. The term is used loosely as a synonym for a robot's morphology, or body plan. For a mobile robot, the choice of form factor is the earliest and most consequential design decision, because it sets the machine's energy budget, the number of [actuators](/wiki/actuator) it needs, its cost, its failure modes, and the range of environments it can enter. Almost every later trade-off, from battery size to control software to unit economics, is downstream of it.

The design space runs, roughly in order of increasing mechanical complexity, from a bolted-down manipulator arm, through a wheeled base carrying an arm, to legged machines such as quadrupeds and bipeds, with several hybrids in between. A recurring engineering claim is that mobility efficiency on flat, prepared ground falls as one moves from wheels toward legs, and that two-legged (bipedal) machines are the most demanding of all to power and balance [1]. That claim is broadly true on hard, level floors, but it is not a universal law: a well-designed biped can match human walking efficiency, and legs buy access to terrain that defeats wheels entirely [4]. The interesting question is not which form factor is "best" but which one a given job actually requires.

## The design space

Mobile-robot morphologies form a spectrum rather than a set of discrete categories, but a working taxonomy runs from fixed manipulators to full [bipedal](/wiki/bipedal_locomotion) humanoids. Each step to the right in the table below tends to add [degrees of freedom](/wiki/degrees_of_freedom), actuators, and cost, and to broaden the terrain the machine can handle. Degree-of-freedom counts below refer to actuated joints and are typical rather than universal; representative machines are illustrative, not exhaustive.

| Form factor | Typical actuated DoF | Mobility domain | Relative build cost | Representative machines |
|---|---|---|---|---|
| Fixed manipulator (bolted-down arm) | 4 to 7 | None (stationary) | Low to medium | FANUC, KUKA, Universal Robots arms |
| Automated guided vehicle (AGV) | 0 mobility (fixed path) | Flat floors, marked routes | Low to medium | Amazon Hercules drive units |
| [Autonomous mobile robot](/wiki/autonomous_mobile_robot) (AMR) | 0 to 2 (plus free navigation) | Flat floors, free routes | Low to medium | Amazon Proteus, MiR, Locus |
| Wheeled base plus arm (mobile manipulator) | 6 to 8 | Flat floors | Medium to high | [Boston Dynamics](/wiki/boston_dynamics) Stretch, Fetch |
| Tracked | 2 plus (optional arm) | Rubble, stairs, soft ground | Medium | EOD and inspection crawlers |
| [Quadruped](/wiki/quadruped_robot) | 12 (3 per leg) | Stairs, curbs, rough terrain | High | [ANYmal](/wiki/anybotics), Spot, Unitree B2 |
| Wheeled-legged hybrid | 12 plus 2 to 4 wheels | Flat floors and rough terrain | High | ANYmal on Wheels, RIVR, Unitree B2-W |
| Wheeled humanoid | 20 to 40 (on a wheeled base) | Flat floors, human spaces | High | Unitree G1-D, various bench platforms |
| Biped humanoid | 20 to 50 plus | Stairs, human-built spaces | Very high | [Digit](/wiki/agility_digit), [Optimus](/wiki/tesla_optimus), Figure |

The pattern is that wheels are cheap, simple, and efficient but require prepared ground, while legs are expensive and power-hungry but negotiate the stairs, curbs, and clutter of environments built for people. Hybrids try to capture both. Where a machine should sit on this spectrum depends less on ambition than on the physics and economics of the specific task, which the rest of this article works through.

## The physics: why legs cost more energy on flat ground

The tool for comparing how much energy different bodies spend to move is the [cost of transport](/wiki/cost_of_transport) (COT), a dimensionless number equal to the energy used per unit weight per unit distance travelled. Formally, COT equals E divided by (m g d), or equivalently power divided by (m g v), where m is mass, g is gravitational acceleration, d is distance, and v is speed [28]. Because the weight term makes the quantity dimensionless, it lets engineers compare a beetle, a bicycle, a truck, and a robot on one axis. The metric traces to Gabrielli and von Karman's 1950 study "What Price Speed?", which plotted the specific power that vehicles need against their speed and produced the diagram now named for them [2], and to the biologist Vance Tucker, whose 1975 paper "The Energetic Cost of Moving About" showed that walking and running are among the least efficient ways for an animal to travel, far behind swimming, flying, and, notably, a human on a bicycle [3].

That last point is the crux of the wheels-versus-legs argument on flat ground: putting a walking animal on wheels drops its cost of transport dramatically. The cleanest robotic demonstration comes from ETH Zurich's ANYmal, a quadruped that can be fitted with actuated wheels on the ends of its legs and can therefore both trot and drive on the same hardware and the same floor. In the 2019 paper "Keep Rollin'", Bjelonic and colleagues report that on flat terrain the wheeled ANYmal "achieves a COT of 0.1 while driving 2 m/s and the mechanical power consumption is 63.64 W", and that this "is lower by 83% w.r.t. the trotting gait and by 17% w.r.t. skating motions" [1]. Driving also broke the platform's previous top-speed record of 1.5 m/s, reaching 4 m/s [1]. Because the driving figure is 83 percent below the trotting figure, the trotting cost of transport implied is near 0.6, which matches the numbers circulated in popular summaries of the result.

Several caveats keep this from being over-read. The 0.1 figure is a mechanical cost of transport, measured on flat terrain, at 2 m/s, for one specific machine (the wheeled-legged ANYmal) in one specific mode; it is not a universal constant for wheels versus legs. Cost of transport is also reported inconsistently across the literature: some papers measure mechanical COT (joint power only) and others electrical or metabolic COT (which includes electronics, heat, and, for animals, basal metabolism), and the two can differ by a large factor [28]. Even for ANYmal, a separate study of hybrid trotting reports different numbers, so cost-of-transport figures from different papers are not directly comparable. The safe reading is directional: on hard, flat, prepared ground, rolling is far cheaper than stepping, for both animals and this robot.

The deeper reason legs cost more is that walking repeatedly raises and lowers the body's mass and accelerates and decelerates the limbs, and naive controllers dissipate that energy at every step. But legs are not doomed to inefficiency. In 2005, Collins, Ruina, Tedrake, and Wisse built bipedal robots based on passive-dynamic walking, machines whose leg swing is governed largely by gravity and inertia rather than by high-gain motors, and showed they could walk with close-to-human efficiency [4][5]. Their Cornell biped reached a mechanical cost of transport of about 0.055, close to the human figure of roughly 0.05, while Honda's Asimo, which walked with stiff, flat-footed, high-gain control, came in around 1.6, more than an order of magnitude higher [4]. On the broader energetic measure that includes electronics and heat, the passive-dynamics robot and a human both sit near 0.2 while Asimo sits near 3.2 [4]. The lesson is that most of a walking robot's inefficiency comes from its control strategy and actuator design, not from the mere fact of having legs, and that biological bipeds, humans included, are not especially cheap movers to begin with, which is exactly Tucker's point [3]. Wheels still win on a warehouse floor; but "two legs is inherently the least efficient way to move" overstates a result that is really about prepared ground and particular designs.

## Stability, actuator count, and cost

Energy is only one axis. Each additional limb and joint is another actuator, another gearbox or [servo motor](/wiki/servo_motor), another controller channel, another sensor to fuse, and another thing that can break. A differential-drive AMR may have two drive motors; a quadruped has twelve coordinated actuated joints; a full humanoid commonly carries between roughly twenty and fifty, before counting the hands [4]. That count drives bill-of-materials cost, control complexity, and reliability in the same direction: up.

Stability follows the same gradient. A wheeled base with three or more contact points is statically stable: switch it off and it stays put. A dynamically balancing machine, whether a two-wheeled self-balancer or a walking biped, is statically unstable and must actively control its balance in real time using [whole-body control](/wiki/whole_body_control), so a software fault or a dead battery can mean a fall, and a fall for a tall humanoid is a high-energy event that can damage the machine or its surroundings. Bipeds are the hardest case because they have the smallest support polygon and the highest center of mass, which is why so much of the humanoid research literature is about balance, foot placement, and fall recovery rather than manipulation.

These costs compound. More actuators mean more mass, which means larger actuators and batteries to move that mass, which means more cost and shorter run time, a spiral familiar to anyone who has designed a legged machine. It is why the price bands in the design-space table climb so steeply from left to right: a wheeled AMR and a research quadruped can differ by an order of magnitude in build cost, and a capable humanoid can differ by another. As of 2026 a Unitree G1 humanoid started around 13,500 to 16,000 US dollars in its base configuration while research and enterprise versions ran far higher [25], Agility's Digit was quoted at upward of 250,000 dollars [15], and Tesla has publicly targeted a sub-30,000-dollar price for [Optimus](/wiki/tesla_optimus) at scale, a figure that remains a target rather than a shipping price [25]. Fixed industrial arms and wheeled AMRs, by contrast, are mature, mass-produced products whose economics are already well understood.

## Where legs actually earn their cost

Legs justify their expense wherever the ground is not prepared for wheels. Stairs are the canonical case: a step is a vertical discontinuity that a wheel cannot climb unless the wheel is larger than the step, whereas a leg simply places a foot on the next tread. Curbs, thresholds, door sills, cables, hoses, debris, mud, sand, and rubble are variations on the same problem. Environments that were built for humans, and legacy sites that predate any thought of robots, are full of exactly these features, which is the core of the argument that machines meant to work in the human world may need to move the way humans do.

Disaster response and industrial inspection are where legged robots have found their firmest footing. Quadrupeds such as [ANYbotics'](/wiki/anybotics) ANYmal and Boston Dynamics' Spot are sold specifically to patrol oil and gas plants, substations, mines, and construction sites, climbing stairs and stepping over pipework that would strand a wheeled cart [23]. The 2015 DARPA Robotics Challenge, which posed disaster-response tasks such as opening doors, turning valves, and climbing over rubble, was one of the first large public tests of legged and semi-legged machines against unstructured terrain.

The accessibility of the built world is quantifiable, and it favors legs less often than intuition suggests. A 2022 study by Harvard's Joint Center for Housing Studies found that only about 3.5 percent of US homes combine the three basic features that let a person with limited mobility, or a wheeled robot, move through them: a no-step (zero-step) entrance, single-floor living, and doorways and halls wide enough to accommodate a wheelchair [6]. The figure comes from the American Housing Survey and applies to the United States only. Read one way, it is an argument for legs: the overwhelming majority of homes have at least one step, so a robot that cannot climb is locked out of most of them. Read another way, it is a caution: even the 3.5 percent that are roll-through-ready still contain stairs to upper floors, and truly step-free operation is rare, so neither wheels nor legs alone solve the home. The statistic is best understood as a measure of how hostile ordinary buildings are to wheels, not as a settled verdict for any one morphology.

## Where wheels dominate

Wherever the floor can be engineered, wheels win decisively, and the clearest example is the modern fulfilment warehouse. [Amazon](/wiki/amazon_robotics) announced in mid-2025 that it had deployed its one-millionth warehouse robot, with the milestone unit delivered to a fulfilment center in Japan, across a network of more than 300 facilities, alongside a generative-AI fleet-coordination model called DeepFleet that it said improved robot travel time by about 10 percent [7][8]. That fleet is entirely wheeled. It includes Hercules drive units that slide beneath and lift inventory pods weighing up to about 1,250 pounds, Pegasus units that carry individual packages on a conveyor deck, Proteus, described as Amazon's first fully autonomous mobile robot, and the Sequoia, Sparrow, and Cardinal systems that store, pick, and place goods at fixed stations [7][8]. Not one of these operational machines has legs.

The reason is economic. Wheeled AMRs are statically stable, need only a handful of motors, are cheap to build and maintain, and can run long shifts and dock themselves to charge, so their uptime and cost per pick are hard for any legged machine to match. They depend, however, on a floor poured to a flatness specification. Concrete floor flatness and levelness are measured with the F-number system defined by ASTM E1155 and adopted into the American Concrete Institute's ACI 117 tolerances, where a higher FF number means fewer local bumps and a higher FL number means a more level slab [19][20]. Narrow-aisle and AGV facilities typically call for something like FF 35 to 50 and FL 25 to 35, and "superflat" floors specified as high as FF 100 and FL 50 are poured for very-narrow-aisle warehouses where lift trucks or robots run at full speed down defined paths, a class of specification that traces to the "superflat" floors developed in the 1970s for very-narrow-aisle trucks [19][20]. Wheeled robots are intolerant of vertical discontinuities: a lip, a gap, or a threshold that a person would step over without noticing can stop a low-clearance drive unit, which is precisely why the floors they run on are engineered so flat in the first place. Legs would buy nothing in that environment and would cost a great deal.

Amazon's wheeled fleet is not the whole of its robotics strategy, and the nuance matters. The company tested Agility Robotics' bipedal Digit at its robotics research and development site south of Seattle beginning in October 2023 [16], and in March 2026 it acquired RIVR, a Zurich company (formerly Swiss-Mile) that builds wheeled-legged quadrupeds able to climb stairs, for last-meter doorstep delivery [17][18]. So the accurate statement is narrower than the slogan: Amazon's operational fulfilment fleet is wheeled because its floors are engineered flat, but the company is actively investing in legged and wheeled-legged machines for the tasks, stairs and doorsteps, where wheels alone fail.

## The wheeled-legged middle ground

Between the pure wheel and the pure leg sits a growing class of hybrids that mount wheels on legs, aiming to drive efficiently on flat ground and walk or step only when the terrain demands it. The [wheeled-legged robot](/wiki/wheeled_legged_robot) is the general name for this class. ANYmal on Wheels, the platform behind the "Keep Rollin'" results, is the research archetype: it drives at a tenth of the cost of trotting on flat ground and lifts its wheeled feet over obstacles when it has to [1]. [Boston Dynamics'](/wiki/boston_dynamics) Handle, first revealed in 2017 as a research machine that combined two wheels with legs and could travel at about 15 km/h and jump around 1.2 meters, explored the same idea for warehouse box-handling [10].

Handle also illustrates why the hybrid is hard to productize. Boston Dynamics moved its warehouse effort from Handle to Stretch, a simpler arm on an omnidirectional wheeled base, and shipped Stretch commercially, with DHL Supply Chain announcing the first commercial deployment in January 2023 [9][27]. The company's engineers were explicit that this was a practicality decision rather than a failure of the concept. Kevin Blankespoor, who led both projects, said that when Handle was tested on truck unloading "it took too long", because "every time Handle grasped a box, it would have to roll back and then get to a place where it could spin itself to face forward and place the box", and that while "Stretch has really taken over our team as far as warehouse products go", Handle is still used "occasionally as a research robot" [9]. The dynamic, balancing hybrid was elegant, but a statically stable arm on wheels did the paying job faster and more reliably.

Outside the warehouse the hybrid is thriving. RIVR (formerly Swiss-Mile), an ETH Zurich spin-off, builds wheeled quadrupeds for doorstep delivery, raised roughly 22 million dollars, ran pilots with Veho and Just Eat Takeaway, and was acquired by Amazon in 2026 [17][18]. Unitree sells the B2-W, a wheeled version of its B2 quadruped that switches between walking and rolling for industrial and rough-terrain use, and LimX Dynamics offers a comparable wheeled quadruped [29]. Ascento builds a two-wheeled, two-legged self-balancing robot with sprung knees for outdoor security patrol, combining the speed of wheels on paths with the ability to hop small obstacles [24]. A parallel "wheeled humanoid" class, a human-like torso and arms mounted on a wheeled or two-wheeled balancing base rather than on walking legs, has become common on Chinese development platforms and includes Unitree's G1-D, launched in late 2025 as a wheeled humanoid for data collection and training. These machines accept that most useful ground is flat while keeping some ability to handle the parts that are not.

## The general-purpose argument for humanoids

The case for building full bipeds, made by the people who actually build them, does not rest on efficiency, where legs lose, but on generality and on the shape of the human-built world. Its strongest form has three parts. First, the world already fits the human body: doorways, stairs, handles, tools, vehicles, and workstations were all designed around human proportions and reach, so a machine with a human shape can in principle use them without anyone rebuilding the environment. NVIDIA's Jensen Huang has put it directly, arguing that "the easiest robot to adapt into the world are humanoid robots because we built the world for us" [21]. Second, one general-purpose body can in theory amortize its cost across many tasks, where a warehouse might otherwise need a dozen specialized machines, an argument Figure's Brett Adcock frames by calling the humanoid "the ultimate deployment vector" for general-purpose AI [22]. Third, humans generate an enormous amount of demonstration data, from video to motion capture, and a robot with roughly human kinematics can absorb that data more directly than a machine with a different body plan, which matters as robot control shifts toward learned [embodied AI](/wiki/embodied_ai) policies [21].

Against these points stand the costs already described: bipeds are the most expensive, most actuator-heavy, least energy-efficient, and hardest-to-balance form factor, and for any single task a specialized wheeled machine will almost always be cheaper and more reliable. The honest version of the humanoid thesis is a bet that falling actuator and compute costs, plus the value of one platform that can be redeployed across many jobs in spaces built for people, will eventually outweigh the per-task penalty, not a claim that the biped is efficient today. Companies including [Agility Robotics](/wiki/agility_robotics), [Figure AI](/wiki/figure_ai), [Apptronik](/wiki/apptronik), Tesla, [Unitree](/wiki/unitree), [UBTECH](/wiki/ubtech), and [AgiBot](/wiki/agibot) are all placing versions of that bet, and it remains genuinely unsettled.

Even among bipeds, the human silhouette is not always copied faithfully, because designers optimize for function over resemblance. Agility's Digit, one of the earliest humanoids in sustained paid warehouse work, has legs modeled on birds rather than humans: a backward-bending lower segment that is functionally an ankle rather than a reversed knee, a design inherited from the ATRIAS and Cassie research platforms that were themselves derived from analyses of how ostriches and other running birds move [26]. Placing the heavy actuators high, near the torso, and keeping the lower leg light reduces the limb's inertia and its cost of transport and improves balance, which is why "bird legs" recur across efficient bipeds [26]. It is a concrete illustration that "humanoid" is a design target to be traded against physics, not a fixed template.

## The 2025 to 2026 scoreboard

The gap between the wheeled and legged worlds is visible in deployment numbers, though the humanoid figures must be quoted carefully because research firms count differently and differ by year and by whether they measure shipments or installations. For 2025, the widely cited estimates of global humanoid volume span roughly 13,000 to 18,000 units.

| Source | 2025 humanoid figure | Notes |
|---|---|---|
| Omdia | About 13,000 units shipped globally [11] | AgiBot first at 5,168 units (39 percent); Unitree about 4,200; UBTECH about 1,000 |
| Counterpoint Research | About 16,000 installations [12] | China roughly 12,800 of the total |
| Yano Research Institute | About 16,580 units shipped [13] | China about 84.8 percent; forecast 7.18 million by 2035 |
| IDC | About 18,000 units [11] | Reported alongside Omdia's lower count |

Two things are consistent across every estimate. First, Chinese manufacturers dominated, accounting for roughly 80 percent or more of 2025 humanoid volume, a point made explicitly by Bloomberg's reporting on the year [14] and reflected in the AgiBot, Unitree, and UBTECH rankings [11]. Second, the absolute numbers are tiny next to the installed base of wheeled machines: the entire world shipped on the order of 13,000 to 18,000 humanoids in 2025, while a single company, Amazon, was operating more than a million wheeled robots [7][8]. The comparison is not perfectly like-for-like, because a warehouse drive unit and a general-purpose humanoid are built for different jobs, but the order-of-magnitude contrast is real and is the empirical core of the wheels-versus-legs debate.

Deployment quality, not just volume, distinguishes the two. Wheeled AMRs run revenue-generating shifts by the hundreds of thousands. Humanoids in 2025 were largely in pilots, demonstrations, and early industrial trials rather than steady production roles, with a handful of genuine paid deployments standing out: Agility's Digit passed a year of continuous operation at a GXO facility and moved more than 100,000 totes there, and also ran paid work with Amazon, Toyota, and Schaeffler [15]. The [humanoid robot market](/wiki/humanoid_robot_market) and its real-world [deployments](/wiki/humanoid_robot_deployments) are growing quickly, but from a base measured in thousands, against a wheeled base measured in millions.

## How companies actually choose

In practice a robotics team does not start from a preferred body plan; it starts from the job and lets the constraints select the form factor. The dominant variables are the environment (how prepared and how cluttered the ground is), the payload, the duty cycle (how many hours per day at what reliability), the terrain (flat, stairs, or unstructured), and the cost target. The table below sketches how those inputs tend to map onto a morphology.

| If the job is | Then the form factor that usually wins |
|---|---|
| Repetitive task at a fixed station, no travel | Fixed manipulator arm |
| Moving goods across an engineered-flat floor, high duty cycle, tight cost | AGV or wheeled AMR |
| Picking or loading at points along a flat route | Wheeled base plus arm (mobile manipulator) |
| Inspection or patrol over stairs, curbs, and pipework | Quadruped |
| Mostly flat routes with occasional steps, curbs, or thresholds | Wheeled-legged hybrid |
| General tasks in unmodified human spaces, many task types, cost secondary | Biped or wheeled humanoid |
| Doorstep or last-meter delivery with steps at the destination | Wheeled-legged quadruped |

The rule of thumb that emerges is to use the simplest form factor the environment allows. If the floor can be engineered flat and the task is narrow, wheels and a fixed arm are almost always cheaper and more reliable. Legs are justified when the environment cannot be changed and contains the steps and clutter that wheels cannot cross, and hybrids are justified when a route is mostly flat but has a few obstacles that would otherwise force a full legged platform. The humanoid is the bet that one machine should handle the widest range of human environments and tasks, accepting a higher cost per task in exchange for generality.

## Open questions

Several questions will decide how the spectrum shakes out over the next decade. Will actuator, battery, and compute costs fall far enough that a general-purpose humanoid becomes cheaper, over its working life, than the several specialized machines it would replace? Will wheeled-legged hybrids capture enough of both worlds to make pure bipeds unnecessary for most logistics and delivery work, or will their added mechanical complexity keep them in a research niche as it did with Handle? How much of the humanoid case survives if warehouses and even homes are progressively redesigned to be robot-friendly, since a flatter, step-free world is one that wheels can serve? And does learned control change the calculus, if human demonstration data really does transfer best to human-shaped bodies, that data advantage could tilt the economics toward the humanoid independently of its mechanical efficiency. None of these is settled as of 2026, and the shipment numbers show a field still early enough that the answers remain open.

## ELI5

Imagine you want a robot to carry boxes. If the floor is smooth and flat, like a big gym, wheels are the best choice: they are cheap, they do not fall over, they need only a couple of motors, and they use very little energy, which is why the giant warehouses have a million robots on wheels and none with legs. But wheels get stuck on stairs, curbs, and door steps. Legs can climb over those, which is why some robots have four legs like a dog or two legs like a person. The catch is that legs are expensive, they have lots of motors, they can tip over, and they use much more energy to walk than wheels use to roll. So the smart move is to match the robot's shape to the place it works: wheels where the ground is smooth, legs where it is bumpy, and sometimes wheels-on-legs to get a little of both. Two legs like a human is the hardest and priciest kind, and people build it anyway because our whole world, doors, stairs, and tools, was made for human bodies.

## See also

- [Robot locomotion](/wiki/robot_locomotion)
- [Cost of transport](/wiki/cost_of_transport)
- [Wheeled-legged robot](/wiki/wheeled_legged_robot)
- [Bipedal locomotion](/wiki/bipedal_locomotion)
- [Quadruped robot](/wiki/quadruped_robot)
- [Autonomous mobile robot](/wiki/autonomous_mobile_robot)
- [Humanoid robot](/wiki/humanoid_robot)
- [Degrees of freedom](/wiki/degrees_of_freedom)

## References

1. Bjelonic, M., Bellicoso, C.D., de Viragh, Y., Sako, D., Tresoldi, F.D., Jenelten, F., Hutter, M. "Keep Rollin': Whole-Body Motion Control and Planning for Wheeled Quadrupedal Robots." IEEE Robotics and Automation Letters, 2019. arXiv:1809.03557. https://arxiv.org/abs/1809.03557
2. Gabrielli, G., von Karman, T. "What Price Speed? Specific Power Required for Propulsion of Vehicles." Mechanical Engineering, vol. 72, no. 10, 1950, pp. 775-781. https://gwern.net/doc/technology/1950-gabrielli.pdf
3. Tucker, V.A. "The Energetic Cost of Moving About." American Scientist, vol. 63, no. 4, 1975, pp. 413-419.
4. Collins, S., Ruina, A., Tedrake, R., Wisse, M. "Efficient Bipedal Robots Based on Passive-Dynamic Walkers." Science, vol. 307, no. 5712, 2005, pp. 1082-1085. https://www.science.org/doi/10.1126/science.1107799
5. Cornell Chronicle. "Robots that simulate life by walking with close-to-human efficiency described by researchers at Cornell, MIT and Delft." February 2005. https://news.cornell.edu/stories/2005/02/robots-walk-close-human-efficiency
6. Scheckler, S., Molinsky, J. "How Well Does the Housing Stock Meet Accessibility Needs?" Harvard Joint Center for Housing Studies, 2022. https://www.jchs.harvard.edu/sites/default/files/research/files/harvard_jchs_housing_stock_accessibility_scheckler_2022_0.pdf
7. Amazon. "Amazon launches a new AI foundation model to power its robotic fleet and deploys its 1 millionth robot." About Amazon, 2025. https://www.aboutamazon.com/news/operations/amazon-million-robots-ai-foundation-model
8. Bishop, T., et al. "Amazon's robot workforce hits 1 million: Here's what they all do." GeekWire, July 2025. https://www.geekwire.com/2025/amazons-robot-workforce-hits-1-million-heres-what-they-all-do/
9. Ackerman, E. "Stretch Is Boston Dynamics' Take on a Practical Mobile Manipulator for Warehouses." IEEE Spectrum, 2021. https://spectrum.ieee.org/stretch-is-boston-dynamics-take-on-a-practical-mobile-manipulator-for-warehouses
10. ROBOTS Guide. "Handle." IEEE / Robots: Your Guide to the World of Robotics. https://robotsguide.com/robots/handle
11. TechNode. "China's AgiBot leads global humanoid robot shipments in 2025, Omdia says." January 2026. https://technode.com/2026/01/09/chinas-agibot-leads-global-humanoid-robot-shipments-in-2025-omdia-says/
12. Counterpoint Research. "Global Humanoid Robot Installations Reach 16,000 Units in 2025 as Mass Production Picks Up Pace." 2026. https://counterpointresearch.com/en/insights/Global-Humanoid-Robot-Installations-Reach-16,000-Units-in-2025-as-Mass-Production-Picks-Pace
13. Yano Research Institute. "Global Shipment of Humanoid Robots Forecasted to Reach 7.18 Million Units by 2035." 2025. https://www.yanoresearch.com/press/press.php/4111
14. Bloomberg. "Chinese Firms Dominated Global Humanoid Robot Shipments in 2025." January 8, 2026. https://www.bloomberg.com/news/articles/2026-01-08/chinese-firms-dominated-global-humanoid-robot-shipments-in-2025
15. Agility Robotics. "Digit Moves Over 100,000 Totes in Commercial Deployment." 2025. https://www.agilityrobotics.com/content/digit-moves-over-100k-totes
16. Heater, B. "Amazon begins testing Agility's Digit robot for warehouse work." TechCrunch, October 2023. https://techcrunch.com/2023/10/18/amazon-begins-testing-agilitys-digit-robot-for-warehouse-work
17. The Robot Report. "Amazon acquires robotic doorstep delivery provider RIVR." March 2026. https://www.therobotreport.com/amazon-acquires-robotic-doorstep-delivery-provider-rivr/
18. The Robot Report. "Swiss-Mile rebrands to RIVR, continues developing wheeled quadrupeds." 2025. https://www.therobotreport.com/swiss-mile-rebrands-to-rivr-continues-developing-wheeled-quadrupeds/
19. Ace Avant. "ASTM E1155: F-Number System (FF/FL Explained)." https://www.aceavant.com/astm-e1155-f-number-system-ff-fl-explained/
20. Archtoolbox. "FF and FL Numbers: Floor Flatness and Levelness." https://www.archtoolbox.com/ff-and-fl-numbers/
21. Kolodny, L. "Nvidia CEO Jensen Huang explains why he's all in on humanoid robotics." CNBC, March 2024. https://www.cnbc.com/2024/03/20/nvidia-ceo-jensen-huang-explains-why-hes-all-in-on-humanoid-robotics.html
22. Adcock, B. (Figure AI). "Figure AI founder Brett Adcock on humanoid robots as the deployment vector for AGI." AOL / TechCrunch, 2024. https://www.aol.com/figure-ai-founder-brett-adcock-190052477.html
23. ROBOTS Guide / ETH Zurich Robotic Systems Lab. "ANYmal." https://robotsguide.com/robots/anymal
24. The Robot Report. "Ascento launches nimble Guard robot following a new $4.3M funding round." https://www.therobotreport.com/ascento-launches-nimble-guard-robot-following-a-new-4-3m-funding-round/
25. RoboZaps. "Unitree G1 Price 2026: From $13,500, EDU Cost and Full Specs" and related pricing coverage. https://blog.robozaps.com/b/unitree-g1-review
26. "Birds Make Better Bipedal Bots Than Humans Do." Scientific American, and Machine Design, "Energy-Efficient Legged Robot Runs Like a Bird." https://www.scientificamerican.com/article/birds-make-better-bipedal-bots-than-humans-do/
27. DHL. "DHL Supply Chain Achieves First Commercial Deployment of Boston Dynamics' Stretch Robot to Unload Trailers and Containers." January 2023. https://www.dhl.com/us-en/home/press/press-archive/2023/dhl-supply-chain-achieves-first-commercial-deployment-of-boston-dynamics-stretch-robot.html
28. "Cost of transport." Wikipedia. https://en.wikipedia.org/wiki/Cost_of_transport
29. RoboStore / The Robot Report. "Unitree B2-W: wheeled quadruped robot." https://robostore.com/products/unitree-b2-w-industrial-quadruped-robotic-dog-with-wheels