Robotics

Robotics is the interdisciplinary study and engineering of robots: embodied machines whose mechanisms, sensors, computation, and actuators are organized to perform tasks in the physical world. It draws on mechanical and electrical engineering, computer science, materials science, control theory, mathematics, and human factors. Its central problems include designing a body that can produce useful motion and force, estimating the state of the robot and its surroundings, deciding what to do, controlling action under physical constraints, and establishing that the complete system performs safely and reliably. ISO 8373 supplies an international vocabulary for the field, while Modern Robotics organizes technical foundations around mechanics, planning, control, manipulation, and mobility ^[1][2].

Robotics is not synonymous with artificial intelligence. A robot may repeat a taught trajectory, follow a feedback controller, accept direct human commands, use a learned policy, or combine all of these. Conversely, most AI systems do not have a physical body. Modern robotics often uses machine learning for perception or decision-making, but mechanics, timing, power, sensing, and safety remain decisive even when a learned model supplies part of the behavior.

The field ranges from fixed manipulators in controlled work cells to mobile machines in changing environments. "Industrial," "service," "medical," "social," and "humanoid" identify different combinations of use, setting, interaction, or body form rather than successive levels of intelligence. Autonomy is likewise task-specific: a system may navigate independently while requiring a person to select goals, approve contact, or handle failures.

Since about 2022 the research center of gravity has moved toward policies learned from large datasets, including vision-language-action models that map camera images and text instructions to robot commands. Industrial practice has moved much more slowly. The International Federation of Robotics (IFR) counted 542,000 industrial robot installations worldwide in 2024 and an operational stock of 4,664,000 units, almost all of them conventionally programmed machines rather than learned systems ^[29]. The distance between what has been shown in a laboratory or a promotional video and what runs unattended in production is the single most useful thing to hold in mind when reading a robotics claim.

Scope and definitions

There is no single informal test that cleanly separates every robot from every automated machine. ISO 8373 supplies a controlled vocabulary for robotics, but everyday and regulatory usage can be narrower or broader than a technical standard ^[1]. For this article, a robotic system has four connected properties:

1. A physical mechanism can change its own configuration or exert force on the environment.
2. Sensors or operator inputs provide information about the robot, the task, or the surroundings.
3. A control system maps goals and observations to commands.
4. The machine can execute a programmed sequence or behavior, or translate operator commands into coordinated motion, without the operator directly supplying the mechanical work.

These properties admit fixed arms, mobile platforms, legged machines, teleoperated or remotely supervised systems, and many task-specific devices. They do not require human-like appearance, conversational ability, or a particular AI method. A numerical-control machine or automated conveyor may share components with a robot, and the boundary depends on reprogrammability, task variability, and the vocabulary used in the relevant industry.

The word "autonomous" needs a stated task and operating context. A robot can be independent in route following but dependent on a person for destination selection, exception handling, maintenance, or safety authorization. NIST's Autonomy Levels for Unmanned Systems work treated mission complexity, environmental difficulty, and human independence as separate dimensions, illustrating why one context-free autonomy score is usually inadequate ^[8].

Robotics also needs to be distinguished from Robotic Process Automation, which automates software workflows and does not require a physical machine. The neighboring field of Embodied AI studies how an agent's body and its coupling to an environment shape intelligent behavior. It overlaps strongly with robotics, but robotics also includes machines that use little or no AI, while embodied-intelligence research can include biological or simulated agents ^[14]. The related marketing term physical AI, popularized by chip and simulation vendors from 2024 onward, groups robotics with autonomous vehicles and industrial simulation; it identifies a commercial category rather than a technical distinction.

System composition

A robot is a system rather than a single algorithm or component. Its design normally couples the following elements ^[2][3]:

Scroll sideways for more →

The mechanical structure may be a chain of rigid links, a mobile chassis, a flying body, a compliant mechanism, or a combination. Its independent configuration variables are its degrees of freedom. More degrees of freedom can enlarge the set of reachable poses or allow obstacle avoidance, but they also add sensing, actuation, planning, and control demands. Mechanical stops, joint limits, payload, center of mass, traction, and structural flexibility bound what software can command.

An actuator may be electric, hydraulic, pneumatic, thermal, piezoelectric, or based on another transduction mechanism. A transmission trades speed, torque, backdrivability, precision, and efficiency. Power electronics, batteries or external supplies, thermal management, and brakes can be as important as the motor itself. An arm usually interacts through an end effector, such as a gripper, suction tool, surgical instrument, sensor, or process tool. The end effector and workpiece are part of the contact problem even when they are sold separately from the arm.

Sensors are often divided into proprioceptive and exteroceptive measurements. Encoders, current sensors, inertial units, and joint force or torque measurements describe the robot's own motion and loading. Cameras, LiDAR, radar, microphones, proximity sensors, and tactile sensing describe the surroundings or contact. The distinction is useful but not absolute: an external camera can measure robot pose, and motor current can provide an indirect estimate of contact force.

Robotic computation is commonly distributed. Fast device and control loops may run on microcontrollers or real-time processors, while perception, planning, and supervision run on more capable computers. Communication delay, clock drift, dropped messages, and stale sensor data can change closed-loop behavior, so software modularity does not remove timing requirements. Robot Operating System is one middleware family; its original description explicitly characterized ROS as a communications and software framework above host operating systems, not an operating system in the traditional process-scheduling sense ^[6].

Hardware and mechatronics

Software attention has shifted faster than hardware, but the mechanical and electrical choices below still set the ceiling on what any policy can do. A learned controller cannot produce a compliant contact on a joint that cannot be backdriven, and it cannot run at 50 Hz on a computer that delivers one inference per second.

Actuation and transmission

Hydraulic actuation offers high force density and tolerates impact well, which is why it dominated early dynamic legged machines. It also requires pumps, hoses, valves, fluid, and thermal management, and it leaks. The clearest recent marker of the shift away from it is Boston Dynamics retiring the hydraulic Atlas research platform in April 2024 and replacing it with an all-electric successor ^[30]. Electric drives now dominate manipulators, mobile bases, and legged machines outside heavy construction and aerospace applications.

Within electric actuation, the transmission ratio is the central design decision. A high-ratio gearbox such as a harmonic drive delivers high torque density and low backlash, which suits precise position control, but reflected inertia and friction scale with the square of the ratio, so the joint resists being pushed. A quasi-direct-drive design pairs a high-torque-density motor with a low reduction, typically single-digit, so the joint stays backdrivable and joint torque can be estimated from motor current without a dedicated torque sensor. Wensing and colleagues formalized this as proprioceptive actuation in their analysis of the MIT Cheetah, and introduced an impact mitigation factor to quantify how much of an impulse reaches the drivetrain rather than being absorbed ^[49]. A series elastic actuator takes a different route, placing a compliant element between motor and load so that force control becomes deflection measurement, at the cost of control bandwidth.

Scroll sideways for more →

Direct force measurement remains common where compliance matters. A wrist-mounted force-torque sensor gives a clean six-axis reading at one location; joint torque sensing gives a distributed picture at higher cost. Current-based estimation is cheap but confounded by friction and gearbox losses, which is exactly the confound a low reduction ratio reduces.

Dexterous manipulation and tactile sensing

Most robots deployed at scale grip with two parallel jaws or a suction cup, because those tools are cheap, reliable, and easy to model. Multi-fingered dexterous hands add degrees of freedom, cost, fragility, and calibration burden in exchange for in-hand reorientation and the ability to use human tools. As of mid-2026 the strongest published demonstrations of learned multi-finger control come from model developers rather than deployed products. Google DeepMind reported on 30 July 2026 that its Gemini Robotics 2 model controlled a 22-degree-of-freedom hand across a set of dexterity tasks, with per-task success rates from 32 percent to 92 percent ^[37]. Those figures come from the developer's own evaluation and describe controlled trials, not fielded use.

Tactile sensing is the sense most obviously missing from robots that otherwise perceive well. Vision-based tactile sensors solved part of the problem by putting a camera behind an elastomer pad and reading its deformation as an image, which converts a hard transduction problem into a computer-vision problem. The GelSight design established this approach and demonstrated geometry, force, and slip estimation at high spatial resolution ^[50]. Meta's FAIR robotics group released Digit 360 in November 2024, an open-sourced fingertip that the company states carries more than 18 sensing features and more than 8 million taxels ^[51]. Those are manufacturer specifications for a research device. Unlike cameras and LiDAR, tactile sensing still has no dominant interface, no standard data format, and no large shared dataset, which is one reason touch appears far less often than vision in learned policies.

Onboard computation

Robot computation splits by timing requirement rather than by capability. Hard real-time loops for current and joint control run on microcontrollers or real-time processors with deterministic scheduling. Perception, mapping, and learned policies run on general-purpose or accelerated computers where a missed deadline degrades performance rather than breaking the machine.

The accelerated tier moved substantially in 2025. NVIDIA announced general availability of the Jetson AGX Thor developer kit on 25 August 2025 at a list price of 3,499 US dollars, with the Jetson T5000 module rated by NVIDIA at 2,070 FP4 teraflops and 128 GB of LPDDR5X memory, which the company states is roughly 7.5 times the AI compute of the previous Jetson AGX Orin generation ^[52]. Vendor throughput figures are peak numbers under favourable precision assumptions and should not be read as delivered policy throughput.

What matters operationally is the rate at which a policy can close its loop, and here the published numbers are sobering. The 55 billion parameter RT-2-X ran at 1 to 3 Hz, and OpenVLA reported roughly 6 Hz on a single consumer GPU ^[33]. Action chunking, in which a policy emits a short sequence of future actions rather than a single step, is the standard way to bridge the gap between a slow policy and a fast machine, and it was central to the ALOHA line of bimanual systems ^[39]. On a battery-powered mobile robot, inference also competes directly with runtime, so edge AI deployment is a power budget question as much as a throughput question.

Modeling and control

Robot models connect geometry, force, and time. A configuration vector, denoted by q, records the variables needed to specify the robot's pose. Forward kinematics maps q to the pose of a body or tool. Inverse kinematics asks for configurations that achieve a desired pose and may have no solution, one solution, or several. The Jacobian relates joint velocities to task-space velocity and helps identify singular configurations, where motion or force capability is lost in some direction ^[2].

For a rigid-link manipulator, one common dynamics form is

$$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + g(q) = \tau + J(q)^T f_{\mathrm{ext}},$$

where M is the mass matrix, the product of C and the configuration velocity represents velocity-dependent terms, g represents gravity, tau is applied joint effort, and the transposed Jacobian maps an external wrench into joint coordinates. Real machines add friction, backlash, elasticity, actuator saturation, contacts, and parameter error. Different bodies require different models, but the principle is general: a controller cannot command arbitrary motion or force without respecting dynamics and constraints.

Feedback control compares a measured or estimated state with a reference and adjusts actuation to reduce error. Position, velocity, torque, force, impedance, and whole-body controllers regulate different quantities. Contact tasks often require compliance or force control because high position accuracy alone can produce excessive force when the environment is uncertain. For systems with coupled constraints, Model Predictive Control can optimize a sequence of actions while repeatedly updating the plan from new state estimates, although its performance depends on the model, objective, constraints, and available computation.

Control layers operate at different time scales. A low-level loop may regulate current or joint torque, a middle layer may track a trajectory or maintain balance, and a supervisor may select tasks and recovery modes. These layers are design patterns rather than universal boundaries. Some learned policies produce low-level commands directly, while other systems keep learning above a conventional safety or tracking controller. On legged and humanoid machines the middle layer is usually a whole-body controller that distributes a desired task-space wrench across joints while respecting contact, friction, and balance constraints.

Perception and state estimation

Sensors do not provide a complete, perfectly synchronized description of the world. Measurements contain noise, bias, occlusion, false detections, limited range, and delay. Computer vision can estimate objects, motion, depth, or scene structure, but a robotic system must relate those estimates to coordinate frames, timing, and actions. Calibration determines transformations among cameras, tools, joints, and the robot base. Poor calibration can turn an accurate image prediction into a failed grasp.

State estimation combines a process model with measurements over time. Sensor fusion can use complementary devices, such as encoders for precise relative motion and external observations for global correction. Probabilistic robotics represents uncertainty explicitly and uses methods such as Bayesian filters, Kalman-family estimators, particle filters, and graph optimization when their assumptions fit the problem ^[3].

Simultaneous Localization and Mapping illustrates the coupling. A mobile robot in an initially unknown environment estimates its pose while constructing or updating a map, yet pose estimation depends on the map and mapping depends on pose. Modern SLAM systems use different sensors and mathematical formulations, and their performance depends on features, motion, calibration, dynamic objects, and the required map representation. A 2016 field review framed robust perception, not merely larger maps, as a continuing challenge and emphasized failures caused by changing or poorly modeled conditions ^[5].

Perception is also active. A robot can move a camera, touch an object, change lighting, or choose a viewpoint to reduce uncertainty. This means sensing and action cannot always be optimized independently. A plan that is geometrically feasible may still be unusable if the robot cannot observe the state needed to execute it.

Planning and autonomy

Planning converts a goal into an executable course of action. Task planning chooses or orders symbolic actions, such as open, pick, move, and place. Motion planning searches for a path through the robot's configuration space while respecting obstacles and kinematic constraints. Trajectory generation adds timing, velocity, acceleration, force, and sometimes dynamic feasibility. A collision-free geometric path is therefore not automatically a safe or controllable trajectory ^[4].

Exact planning becomes difficult as the number of degrees of freedom, moving objects, contacts, and uncertainty increase. Search, sampling-based methods, optimization, model-predictive methods, and reactive behaviors offer different completeness, optimality, speed, and robustness properties. Many practical systems combine a slower global planner with faster local collision avoidance and feedback. When the environment changes outside the planning assumptions, the system needs replanning, a safe stop, operator assistance, or another defined recovery.

The planning problem also depends on what the robot knows. A fully observed deterministic formulation differs from one with uncertain object pose, hidden state, or unpredictable people. Planning under uncertainty may value information-gathering actions as well as progress toward the task ^[3][4]. Stating that a robot "can plan" is incomplete unless the goal representation, horizon, uncertainty model, constraints, and failure policy are identified.

Autonomy is implemented across the entire stack, not in one "brain." Route selection may be autonomous while grasp approval is manual; a vehicle may control speed and steering within a mapped area while a remote operator handles blocked routes. Teleoperation, shared control, supervised autonomy, and task autonomy are different allocations of authority. Their suitability depends on communication delay, operator workload, consequence of error, and the structure of the environment ^[7][8].

Robot learning

Robot learning uses data or experience to acquire perception models, dynamics models, value functions, policies, or other components. Learning can replace a hand-designed module, tune parameters, provide a prior, or operate inside a model-based system. It does not eliminate the need to define observations, actions, objectives, constraints, data collection procedures, and evaluation conditions.

Learning from demonstration and teleoperation at scale

Imitation learning, also called learning from demonstration in much of the robotics literature, derives behavior from examples supplied through teleoperation, kinesthetic teaching, observation, or other interfaces. Demonstrations can make intent easier to specify than an explicit reward, but the learned behavior reflects coverage and quality of the examples. Differences between the demonstrator and robot, compounding errors outside the demonstrated state distribution, and ambiguous intent remain important limitations ^[9].

What changed after 2023 was the cost of collecting demonstrations rather than the theory. The ALOHA family of low-cost bimanual teleoperation rigs, and in particular Mobile ALOHA, showed that a mobile bimanual platform assembled from commodity parts could gather whole-body demonstrations, and that with 50 demonstrations per task, co-training on a static-arm dataset increased success rates by up to 90 percent on the mobile manipulation tasks studied ^[39]. The DROID approach of standardizing hardware across institutions produced the largest openly released manipulation dataset of its kind: 76,000 teleoperated trajectories over roughly 350 hours, spanning 564 scenes and 86 tasks, gathered by 50 operators using 18 identical Franka arms at 13 institutions over 12 months ^[38]. The Universal Manipulation Interface removed the robot from data collection altogether, using a handheld gripper with a wrist-mounted fisheye camera so that demonstrations can be recorded anywhere and later mapped onto arms with different kinematics ^[40].

The scaling behavior of this data is better characterized than it was. Lin and colleagues collected more than 40,000 demonstrations and ran more than 15,000 real robot rollouts to measure how single-task policy generalization varies with training data, and found an approximately power-law relationship with the number of training environments and objects, but strongly diminishing returns from adding demonstrations once a threshold per environment-object pair was reached ^[41]. Their headline configuration, 32 environments each with a distinct object and 50 demonstrations, reached about 90 percent success on new environment and object combinations for the task studied. The practical reading is that diversity buys generalization and repetition mostly buys reliability, which is close to the opposite of the intuition carried over from language-model pretraining.

Architecture work moved in parallel. Action chunking, where a policy predicts a short horizon of future actions at once, reduced compounding error and let slow models drive fast hardware. Diffusion policy formulations model the demonstration distribution as a denoising process, which handles multimodal demonstrations, where several different actions are equally correct, better than a regression that averages them into an unusable middle.

Reinforcement learning and simulation

Reinforcement learning optimizes behavior from reward and interaction. Robotics adds costly samples, continuous state and action spaces, delayed outcomes, hardware wear, and safety constraints. Model-free and model-based methods make different use of prior knowledge and data. Results that succeed for one robot, task, initialization range, and training procedure do not establish reliability in a broader deployment domain ^[10].

Reinforcement learning has been much more successful for locomotion than for manipulation, and the reason is instructive. Locomotion has a reward that is easy to write down (move forward, stay upright, do not exceed torque limits), a state that proprioception largely observes, and contact events that a rigid-body simulator approximates acceptably. Contact-rich manipulation has none of these properties: success is often a discrete event, the relevant state includes object properties the robot cannot see, and simulated contact with deformable or articulated objects diverges quickly from reality.

GPU-parallel simulation removed the sample-cost objection for the locomotion case. Rudin and colleagues trained quadrupedal locomotion policies for the ANYmal platform on a single workstation GPU with thousands of simulated robots in parallel, reporting flat-terrain policies in under four minutes and rough-terrain policies in about twenty ^[47]. Frameworks including Isaac Lab and MuJoCo made this style of training routine. Whether the same trick helps manipulation depends entirely on whether the simulator models the relevant contact, and often it does not.

Simulation and sim-to-real transfer

Simulation can generate data and expose a controller to conditions that are costly or dangerous to stage on hardware. The transfer gap arises because geometry, friction, compliance, sensing, timing, and disturbances are only approximated. Sim-to-real transfer methods address this in different ways. In a 2017 experiment, domain randomization varied simulated visual properties so an object detector and grasping pipeline could transfer to the studied real setup; that result demonstrates one method under one task, not that randomized simulation makes physical validation unnecessary ^[11].

Domain randomization has since split into recognizably different practices. Visual randomization varies textures, lighting, and camera placement so the perception front end does not overfit to rendering artifacts. Dynamics randomization varies mass, friction, motor constants, and latency so the controller learns a policy robust across a family of plausible physics rather than one exact model. A third strand, system identification, moves in the opposite direction and fits simulator parameters to measurements from the specific machine. These are complementary: randomization buys robustness at the cost of conservatism, identification buys accuracy at the cost of generality.

The clearest transfer successes are in legged locomotion. Lee and colleagues trained a blind quadrupedal controller entirely in simulation using a teacher-student scheme, where a privileged teacher policy with access to simulator ground truth supervises a student that sees only proprioceptive history, and deployed it on ANYmal across natural terrain including mud, vegetation, water, and snow ^[48]. The pattern of that result, train with privileged information in simulation and distil into something that runs from real sensors, recurs widely. In manipulation, the equivalent claim remains weaker, and most strong manipulation results still depend substantially on real demonstrations.

Vision-language-action models

Large pretrained models have added another line of research. A vision-language-action model maps visual observations and language instructions to robot actions. RT-2, published at the 2023 Conference on Robot Learning, encoded actions as tokens and evaluated transfer from vision-language pretraining in a particular manipulation setting ^[12]. The 2024 Open X-Embodiment work pooled datasets from multiple institutions and reported positive transfer across several robot embodiments in its experiments ^[13]. These studies support research on cross-task and cross-robot policies, but they do not establish embodiment-independent control or a universally capable robot.

The design space narrowed quickly after RT-2. OpenVLA, released in June 2024, put a 7 billion parameter open-weights model on top of a Llama 2 backbone with fused DINOv2 and SigLIP visual features, trained on 970,000 real robot episodes drawn from Open X-Embodiment, and reported outperforming the 55 billion parameter RT-2-X by 16.5 percentage points of absolute task success across 29 tasks and several embodiments ^[33]. That comparison is between a small model trained on the open dataset and a much larger closed model, so it says as much about the value of the data pooling as about parameter count.

Physical Intelligence took a different action representation, attaching a flow-matching action expert to a pretrained vision-language backbone so the policy emits continuous action chunks rather than discretized tokens. Its pi-0.5 model, published in April 2025, co-trained on heterogeneous sources including multiple robot platforms, web data, high-level semantic subtask prediction, and object detections, and reported that a mobile manipulator could clean kitchens and bedrooms in homes absent from the training data, with episodes running roughly 10 to 15 minutes ^[34]. The company's successor post in April 2026 described pi-0.7 as a steerable model prompted with speed, quality, and modality metadata as well as language, and reported normalized throughput between 0.9 and 1.5 times that of task-specific fine-tuned baselines on laundry folding, while acknowledging that the team could not attribute specific capabilities to specific training episodes ^[35]. Both are company publications describing the company's own evaluations.

Google DeepMind's Gemini Robotics line separated the problem into two models: an embodied-reasoning model that handles spatial understanding, planning, and progress estimation, and a visuomotor VLA that produces actions. The 1.5 release in September 2025 added a motion transfer mechanism intended to let skills demonstrated on one embodiment appear on another, interleaved actions with an internal reasoning process expressed in natural language so the robot "thinks before acting," and claimed a new state of the art for embodied reasoning on visual and spatial understanding, task planning, and progress estimation ^[36]. Gemini Robotics 2, announced 30 July 2026, extended control from tabletop manipulation to whole-body humanoid control and reported adaptation to new embodiments with a few hours of data and typically fewer than 200 examples, alongside per-task success rates published as bar charts rather than in a peer-reviewed protocol ^[37]. NVIDIA's GR00T line and open community models such as SmolVLA pursued the same architecture at smaller scale and with open weights.

Cross-embodiment data and transfer

The Open X-Embodiment collaboration is the reference point for cross-embodiment work. It consolidated 60 existing datasets from 21 institutions into more than one million real robot trajectories covering 22 distinct embodiments and 527 skills ^[13]. The reported results were positive but bounded: RT-1-X achieved roughly a 50 percent higher success rate than the original methods contributed by the collaborating institutions on the smaller-scale domains, and the larger vision-language-model-based RT-2-X showed about a threefold improvement on generalization evaluations relative to a model trained only on data from the evaluation embodiment. Transfer was not uniform, and the large-data domains benefited least, which is what one would expect if pooling mostly helps where local data is scarce.

Two harder problems persist. Action spaces differ across robots in dimension, frame, and units, so pooled datasets must either normalize into a lowest common denominator (usually end-effector pose plus gripper state) or carry embodiment tokens that the model learns to condition on. Both choices discard information. Observation spaces differ too: camera placement, field of view, and the presence or absence of wrist cameras vary by platform, and a 2026 survey of VLA datasets and benchmarks argued that these fidelity and composition trade-offs, rather than model architecture, now constrain progress most ^[62].

How well do learned policies generalize?

Not nearly as well as the language-model analogy suggests. This is the point where robotics claims most often outrun their evidence, and the recent evaluation literature is unusually direct about it.

LIBERO-Plus applied controlled perturbations along seven dimensions to state-of-the-art VLA models evaluated on the LIBERO benchmark, and reported "consistent brittleness beneath apparent competence": success rates fell from around 95 percent to below 30 percent under modest changes in camera viewpoint or robot initial state, while the same models proved largely unresponsive to changes in the language instruction, with evidence that they were ignoring it ^[43]. A policy that degrades catastrophically when a camera is nudged and does not notice when the instruction changes has not learned the task in the sense the word usually implies.

Three observations follow, and they are worth stating plainly:

1. Generalization in robot learning has so far been demonstrated over objects and scene layouts within a training distribution, not over embodiments, sensor configurations, or task structure. Cross-embodiment transfer works best when the target embodiment resembles something already in the pool.
2. Data diversity, not data volume, drives what generalization there is ^[41]. Collecting a thousand more demonstrations in one kitchen adds much less than fifty demonstrations in a new kitchen.
3. A high benchmark score does not establish that a policy uses all of its inputs. Instruction following in particular has to be tested by varying the instruction, which is how LIBERO-Plus found that the models it examined were largely unresponsive to language ^[43].

The phrase "robot foundation model" is therefore an architectural or training aspiration, not a certification of generality. Models built from Foundation Models still depend on action representation, robot geometry, sensors, control frequency, adaptation data, and safety layers. Real-world policy comparison remains difficult because tasks and laboratories differ; the 2025 RoboArena study proposed distributed, double-blind pairwise trials precisely because fixed centralized tests do not capture all settings ^[23].

Foundation models for robotics

The design question is not whether a pretrained multimodal model can drive a robot, but which parts of the stack it should replace and which it should leave alone. Practice has converged on a division of labor that is fairly stable across otherwise different systems.

A pretrained model is good at things that require world knowledge and are tolerant of latency: recognizing that a mug is a mug and that it belongs in the dishwasher, decomposing "clean the counter" into an ordered list of subtasks, noticing that a step failed, and deciding to ask for help. Classical control is good at things that require guaranteed timing and respect for physics: keeping a leg under the center of mass, limiting contact force, staying inside joint limits, tracking a trajectory at kilohertz rates, and stopping when a safety input opens. Neither is a good substitute for the other.

The arc from SayCan to current systems is the clearest illustration. SayCan, published in 2022, used a language model to propose candidate skills and a set of learned value functions to score whether each skill was feasible from the current state, so the language model supplied semantics and the robot supplied grounding. On 101 tasks in a mock kitchen, the PaLM-based version selected the correct skill sequence 84 percent of the time and executed successfully 74 percent of the time ^[53]. End-to-end VLAs then collapsed this into one network, which improved low-level dexterity but lost the explicit interface where a plan could be inspected. Current systems have partly re-introduced the split: pi-0.5 predicts a semantic subtask before predicting actions ^[34], and the Gemini Robotics line ships the reasoning model and the action model as separate artifacts ^[36].

A useful way to read any such system is by update rate, because the rate separation is what makes the architecture work:

Scroll sideways for more →

Three consequences of this structure are easy to miss. First, the learned layer inherits the physical limits of the layer beneath it: a VLA cannot command a force the drive cannot produce or a compliance the gearbox will not allow. Second, the safety layer is deliberately not learned, which is why a learned policy failing usually looks like a stalled or clumsy robot rather than a dangerous one, provided the layering is honest. Third, latency accumulates across layers, so a policy that runs at 5 Hz in a laboratory with a tethered workstation may behave differently when moved onto a battery-powered Jetson module.

What foundation models have not changed is equally worth stating. Calibration still determines whether a correct visual prediction becomes a successful grasp. Gripper design still determines which objects are graspable at all. Cycle time, mean time between interventions, and recovery behavior still determine whether a system is economically usable. A pretrained model that raises task success from 40 percent to 70 percent is a large research result and still an unusable product in a process that needs 99.9 percent.

Human interaction

Human-Robot Interaction studies how people and robots communicate, coordinate, share control, and understand one another. It includes physical interaction, user interfaces, team organization, training, social cues, and the effects of automation on human attention and workload. A classic survey organized the field around the degree of autonomy, the nature of information exchange, team structure, and the physical or temporal separation between people and robots ^[7].

Interaction quality is not reducible to speech or a human-like face. A clear status light, predictable path, well-designed hand-guiding mode, or timely request for help can matter more than natural-language fluency. Evaluation should include whether users can form an accurate mental model of capability, detect faults, recover from errors, and maintain appropriate attention. Trust is useful only when calibrated to actual performance.

A collaborative robot is not inherently safe merely because it is compact, rounded, or marketed for collaboration. ISO guidance treats collaborative operation as an application in which a robot system and a person can share a workspace under specified safeguards. ISO/TS 15066 supplements industrial-robot safety requirements with methods such as safety-rated monitored stop, hand guiding, speed and separation monitoring, and power and force limiting ^[17]. Risk depends on the complete application, including tool, payload, speed, layout, sensing, and foreseeable contact.

The word "robotic" also does not imply autonomy. Many surgical robots are surgeon-controlled computer-assisted systems. The US Food and Drug Administration explicitly notes that robotically assisted surgical devices under its description cannot perform surgery without direct human control ^[20]. This is an important reminder to state who selects goals, who authorizes actions, and what the system does when communication or perception fails.

Language interfaces change the failure modes rather than removing them. A robot that accepts free-form instructions must decide what an ambiguous instruction means, and the evidence that current policies partly ignore instruction text ^[43] means an operator can receive confident-looking motion that does not correspond to what was asked. Interfaces that expose the system's interpretation before execution, rather than only its motion during execution, remain more useful than fluency.

Robot classes and applications

Robot categories use several independent axes:

Scroll sideways for more →

An industrial robot is defined by an industrial context and application, not simply by size or precision. A service robot performs useful tasks for people or equipment outside industrial automation, but that umbrella covers professional and personal systems with very different hazards. ISO 13482:2014, for example, addresses specified personal-care robot types and excludes industrial, medical, military, airborne, and waterborne robots from its scope ^[18].

Mobility labels likewise do not determine intelligence. An Autonomous Mobile Robot may navigate among changing obstacles, while an automated guided vehicle may follow a more constrained route. Autonomous vehicles and drones are mobile robotic systems when considered as embodied sensing, planning, and control platforms, but each has domain-specific regulation and engineering requirements.

Morphology labels should not be treated as capability rankings. A Humanoid Robot resembles aspects of the human body, while a Social Robot is defined by interaction goals. Either can be teleoperated, scripted, or partly autonomous. A purpose-built arm or wheeled platform can outperform a human-shaped machine on a task for which its body and environment were designed. Legged and bipedal locomotion buys the ability to cross terrain that defeats wheels, at a cost in energy, control complexity, and reliability that is only worth paying where the terrain actually demands it.

Medical robotics illustrates why application scope matters. Robotically assisted surgical equipment has a dedicated IEC safety and essential-performance standard, IEC 80601-2-77, separate from industrial and personal-care standards ^[19]. Rehabilitation devices, prostheses, laboratory automation, and surgical assistance can fall under different definitions and regulatory pathways even when they share actuators or control methods.

Evaluation, safety, and deployment

Measuring performance

Robot performance must be measured against an intended task and operating domain. Useful measures can include pose accuracy and repeatability, path deviation, mapping accuracy, force tracking, task completion, cycle time, energy use, intervention rate, recovery success, and human workload. ISO 18646-2:2024, for example, defines navigation tests for specified mobile service robots using measures such as pose accuracy, obstacle avoidance, path deviation, narrow-passage behavior, and mapping accuracy, while explicitly separating those tests from safety validation ^[21].

Component scores do not automatically predict system success. A grasp requires perception, calibration, planning, control, tooling, and object mechanics to work together. NIST's robotic-system performance framework therefore decomposes assembly tasks into perception, mobility, dexterity, and safety capabilities and then studies how those measurements compose at system level ^[22]. Testing should report hardware, software version, payload, environment, initial conditions, operator assistance, number of trials, and failure definitions.

Reproducibility of robot learning results

Robot learning results are harder to reproduce than results in most other machine-learning subfields, and the reasons are structural rather than cultural:

Scroll sideways for more →

The trial-count problem has been quantified. A 2026 benchmark analysis observed that real-robot VLA evaluations typically rest on binary success at a fixed timeout with 25 or fewer rollouts per condition, almost always without confidence intervals or paired statistical comparison, and argued that such cohorts cannot reliably resolve close comparisons ^[45]. That work proposed a distributional alternative: use the time-to-success cumulative distribution as the primitive, score with a throughput measure anchored to same-fixture human teleoperation and reported with bootstrap confidence intervals, and test significance separately. On four publicly available VLAs it resolved two close comparisons that binary metrics did not, left the closest pair unresolved within its trial budget, and found the best evaluated model roughly seven times slower per operation than the human reference ^[45].

Several efforts attack the cost side of the same problem. SIMPLER built simulated counterparts of common real setups and, across more than 1,500 paired simulated and real evaluations spanning two embodiments and eight task families, reported strong correlation between simulated and real policy ranking, using mean maximum rank violation and Pearson correlation as its quality measures ^[44]. SureSim formalized the mixture of large-scale simulation with small-scale hardware testing as prediction-powered inference, correcting simulator bias with paired evaluations and producing confidence intervals on real-world performance, reporting 20 to 25 percent savings in hardware evaluation effort at comparable bound tightness ^[46]. NVIDIA's Isaac Lab-Arena packages GPU-parallel policy evaluation with a composable task library, with the company reporting up to 13.5 times faster evaluation than sequential execution ^[61].

Standard benchmarks help only when their protocol is respected. LIBERO provides 130 simulated manipulation tasks in four suites that isolate shifts in object type, spatial arrangement, goal, and combinations of the three, with human teleoperated demonstrations for each ^[42]. It is a knowledge-transfer benchmark, not a general competence test, and reporting a LIBERO number without the suite, the demonstration budget, and the seed count communicates very little.

Deployment at scale

Installed-base statistics are the best available corrective to demonstration-driven impressions of the field. The picture as of the most recent IFR figures is that robots work at very large scale in a small number of highly structured settings, and barely at all elsewhere.

Scroll sideways for more →

Geography is concentrated as well. Asia accounted for 74 percent of 2024 industrial installations against 16 percent in Europe and 9 percent in the Americas, with China alone installing 295,000 units and holding an operational stock of 2,027,000 ^[29]. In the United States, the IFR reported at the Automate trade show in June 2026 that 2025 installations rose about 11 percent to roughly 38,000 units, the first increase after two consecutive annual declines, with more than a third going to automotive and a 41 percent rise in non-manufacturing sectors such as warehousing and logistics ^[32].

The common feature of the domains where robots work at scale is that someone engineered the environment as well as the machine. Warehouse robots run on flat floors with fiducial markers and a fleet manager that owns routing. Industrial arms work in cells where part presentation is fixed. Milking robots operate on cows that voluntarily enter a stall built around the machine. Surgical systems are teleoperated by a trained surgeon ^[20]. Where the environment cannot be engineered, deployment thins out quickly.

The domains where demonstrations most clearly outpace deployment are general-purpose home assistance, unstructured outdoor manipulation, and general-purpose humanoids. The IFR's own August 2025 position paper on humanoid robots concluded that humanoids will complement rather than replace existing robot types and that mass adoption as universal household helpers may not happen in the near or medium term ^[31]. A reader assessing any specific claim in these areas should ask a short list of questions: how many distinct physical sites, how many hours of unattended operation, what intervention rate, and whether the evidence is an audited deployment or an edited video produced by the vendor. Where a capability rests only on a company video, that is what it rests on.

Safety engineering and standards

Safety begins with the intended use, reasonably foreseeable misuse, hazards, and acceptable risk. Measures can include mechanical limits, guarded space, safe speed and separation, force limiting, redundant sensing, brakes, emergency stops, fault detection, degraded modes, and operator procedures. Cybersecurity and communication failures matter when commands or safety-relevant data cross a network. No single list applies to every robot.

Industrial standards show the application-specific structure. ISO 10218-1:2025 covers safety requirements for industrial robots as partly completed machinery ^[15]. ISO 10218-2:2025 covers integration of industrial robot applications and cells, including commissioning, operation, maintenance, and decommissioning ^[16]. Collaborative applications add the concerns addressed by ISO/TS 15066 ^[17]. Personal-care and medical systems use different standards because users, environments, contact, and consequences differ ^[18][19].

The 2025 revision of ISO 10218 was the first substantial rework since 2011 and changed the shape of the collaborative-robot literature. Most of the requirements previously carried by ISO/TS 15066:2016, including the power and force limiting method, were absorbed into ISO 10218-2:2025, and the revision made functional-safety requirements explicit where they had been implied, added cybersecurity requirements, and introduced robot classifications with associated functional-safety requirements and test methods ^[60]. The biomechanical limits themselves were not the thing that changed. Documentation written before 2025 that treats ISO/TS 15066 as a free-standing specification is now describing a superseded arrangement, even though the technical specification remains listed in ISO's catalogue.

Mobile robots have their own track. In North America, ANSI/A3 R15.08-1 (2020) sets design requirements for industrial mobile robots and R15.08-2 (2023) covers integration of mobile robot systems and applications, with a third part addressing users ^[55]. A machine that moves through spaces occupied by people cannot be safeguarded by a fence, so the standards lean on speed and separation monitoring, protective stopping, and defined operating zones rather than on physical guarding.

Functional safety underneath all of this is a matter of architecture rather than intent. Safety functions are assigned an integrity target (a performance level under ISO 13849 or a safety integrity level under IEC 61508 and IEC 62061), and the implementation must show, with evidence, that dangerous failure rates and diagnostic coverage meet that target. This is why safety-rated monitored stop is implemented in certified hardware and firmware rather than in the layer that runs the policy.

Learned policies and certification

Learning-enabled systems add distribution shift and model-update questions to established robotic hazards. Evaluation should include perturbations, rare but plausible conditions, sensor degradation, timing faults, contact errors, and recovery behavior. Simulation is useful for coverage, but hardware tests reveal unmodeled dynamics and integration failures. A benchmark success rate is evidence only for its protocol; deployment claims require tests that reflect the actual robot, users, environment, and consequence of error ^[11][23].

The open question is structural. Functional-safety regimes were built for systems whose behavior can be specified, decomposed, and shown to fail at a bounded rate. A learned policy has no specification in that sense, its failure modes are discovered rather than enumerated, and its behavior changes when the model is updated. ISO 10218:2025 does not resolve this; the revision addresses functional safety and cybersecurity but is not an AI standard ^[60]. The most direct standards response so far is ISO/IEC TR 5469:2024, a technical report on functional safety and AI systems that describes AI properties, associated risk factors, applicable methods, and constraints on where AI can sit relative to a safety function ^[54]. A technical report is informative rather than normative, and a technical specification developed from it by a joint working group of ISO/IEC JTC 1/SC 42 and IEC TC 65 is the route by which normative requirements would arrive.

In the meantime, practical systems mostly sidestep certification of the learned component. The common pattern is to keep the learned policy outside the safety function entirely: it proposes motion, and a certified layer beneath it enforces speed, force, workspace, and stopping behavior regardless of what was proposed. This preserves the existing certification argument at the cost of capping what the policy is permitted to do, which is part of why the most capable published manipulation demonstrations run on research hardware rather than in certified cells.

Reliability and operations

Reliability also includes maintenance and operations. Wear changes friction and backlash, batteries age, lenses become dirty, calibrations drift, networks fail, and people adapt their behavior around a machine. Logging, version control, diagnostics, inspection intervals, spare parts, training, and incident review are part of the deployed robotic system, even though they are not visible in a demonstration.

Learned components add a version-control problem that conventional robots do not have. A model update can change behavior across every task the robot performs, including tasks that were not retested, so fleets running learned policies need staged rollout, per-site regression testing, and the ability to roll back. The service-robot market's move toward robot-as-a-service contracting, where IFR reported the rented and subscribed fleet grew 31 percent in 2024, partly reflects this: when the vendor keeps responsibility for the fleet, it also keeps responsibility for what a model update does to it ^[30].

Historical development

The modern word "robot" came from Karel Capek's play R.U.R., published in 1920 and first performed in 1921. Capek credited his brother Josef with suggesting a word derived from the Czech robota, associated with compulsory labor. The play's manufactured workers were biological rather than metal machines, so the word's original fictional use was not a technical definition of today's robot ^[24].

Programmable industrial manipulation emerged from advances in control, computing, and factory automation. George Devol filed his "Programmed Article Transfer" patent in December 1954; the patent was issued in 1961 and described a reprogrammable transfer mechanism ^[25]. That year, the first Unimate used on an assembly line began unloading a die-casting press at a General Motors plant in Trenton, New Jersey. The Henry Ford museum describes it as the first such assembly-line installation and preserves the machine as an industrial artifact ^[26].

Research robotics broadened the focus from repeatable factory motion to perception and reasoning. SRI's Shakey project, conducted from 1966 through 1972, integrated a mobile platform with environmental sensing, route finding, planning, and simple object rearrangement ^[27]. Its importance lies less in modern hardware capability than in demonstrating that perception, world modeling, planning, and action could be studied as one embodied system.

Later work challenged the assumption that intelligent behavior required a complete central world model. Rodney Brooks's 1991 paper "Intelligence without representation" described layered activity-producing systems coupled directly to perception and action ^[28]. Model-based planning, reactive control, probabilistic estimation, and learning have since developed alongside one another. Contemporary robots often combine them: a learned perception model may feed a probabilistic estimator, a planner may propose a trajectory, and a conventional controller may enforce motion and force limits.

The following table summarizes the periods most often used to organize the field, with the caveat that they overlap heavily and that older methods remain in production use.

Scroll sideways for more →

Robotics remains a systems discipline. Progress in one component matters only when the body, sensors, computation, control, human role, and operating conditions support the complete task. That is why a useful robotics claim specifies not just what a machine did, but which robot did it, under what conditions, with what assistance, how often, and how failure was measured.

References

1. ^International Organization for Standardization. *ISO 8373:2021 Robotics - Vocabulary.* 2021. iso.org/...75539
2. ^Lynch, K. M., and Park, F. C. *Modern Robotics: Mechanics, Planning, and Control.* Cambridge University Press, 2017. modernrobotics.northwestern.edu/nu-gm-book-resource
3. ^Thrun, S., Burgard, W., and Fox, D. *Probabilistic Robotics.* MIT Press, 2005. mitpress.mit.edu/...probabilistic-robotics
4. ^LaValle, S. M. *Planning Algorithms.* Cambridge University Press, 2006. lavalle.pl/planning
5. ^Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I., and Leonard, J. J. "Past, Present, and Future of Simultaneous Localization and Mapping: Toward the Robust-Perception Age." *IEEE Transactions on Robotics* 32, no. 6 (2016): 1309-1332. arxiv.org/...1606.05830
6. ^Quigley, M., Conley, K., Gerkey, B., Faust, J., Foote, T., Leibs, J., Berger, E., Wheeler, R., and Ng, A. "ROS: an Open-Source Robot Operating System." *ICRA Workshop on Open Source Software*, 2009. robotics.stanford.edu/...icra2009-ros.pdf
7. ^Goodrich, M. A., and Schultz, A. C. "Human-Robot Interaction: A Survey." *Foundations and Trends in Human-Computer Interaction* 1, no. 3 (2007): 203-275. scholarsarchive.byu.edu/...940
8. ^Huang, H., Pavek, K., Novak, B., Albus, J. S., and Messina, E. R. "A Framework for Autonomy Levels for Unmanned Systems (ALFUS)." NIST, 2006. nist.gov/...autonomy-levels-unmanned-systems-alfus
9. ^Argall, B. D., Chernova, S., Veloso, M., and Browning, B. "A Survey of Robot Learning from Demonstration." *Robotics and Autonomous Systems* 57, no. 5 (2009): 469-483. eecs.northwestern.edu/...09ras_argall.pdf
10. ^Kober, J., Bagnell, J. A., and Peters, J. "Reinforcement Learning in Robotics: A Survey." *The International Journal of Robotics Research* 32, no. 11 (2013): 1238-1274. journals.sagepub.com/...0278364913495721
11. ^Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., and Abbeel, P. "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World." *2017 IEEE/RSJ International Conference on Intelligent Robots and Systems*, 23-30. arxiv.org/...1703.06907
12. ^Zitkovich, B., et al. "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control." *Proceedings of the 7th Conference on Robot Learning*, PMLR 229 (2023): 2165-2183. proceedings.mlr.press/...zitkovich23a
13. ^Open X-Embodiment Collaboration. "Open X-Embodiment: Robotic Learning Datasets and RT-X Models." *2024 IEEE International Conference on Robotics and Automation*, 2024. robotics-transformer-x.github.io
14. ^Pfeifer, R., and Bongard, J. *How the Body Shapes the Way We Think: A New View of Intelligence.* MIT Press, 2006. mitpress.mit.edu/...e-body-shapes-the-way-we-think
15. ^International Organization for Standardization. *ISO 10218-1:2025 Robotics - Safety requirements - Part 1: Industrial robots.* 2025. iso.org/...73933
16. ^International Organization for Standardization. *ISO 10218-2:2025 Robotics - Safety requirements - Part 2: Industrial robot applications and robot cells.* 2025. iso.org/...73934
17. ^International Organization for Standardization. *ISO/TS 15066:2016 Robots and robotic devices - Collaborative robots.* 2016. iso.org/...62996
18. ^International Organization for Standardization. *ISO 13482:2014 Robots and robotic devices - Safety requirements for personal care robots.* 2014. iso.org/...53820
19. ^International Electrotechnical Commission. *IEC 80601-2-77:2019 Medical electrical equipment - Part 2-77: Particular requirements for the basic safety and essential performance of robotically assisted surgical equipment*, including Amendment 1:2023. webstore.iec.ch/...29933
20. ^US Food and Drug Administration. "Computer-Assisted Surgical Systems." fda.gov/...computer-assisted-surgical-systems
21. ^International Organization for Standardization. *ISO 18646-2:2024 Robotics - Performance criteria and related test methods for service robots - Part 2: Navigation.* 2024. iso.org/...82643
22. ^National Institute of Standards and Technology. "Performance Assessment Framework for Robotic Systems." Updated 2025. nist.gov/...e-assessment-framework-robotic-systems
23. ^Atreya, P., et al. "RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies." *Proceedings of the 9th Conference on Robot Learning*, PMLR 305 (2025): 336-364. proceedings.mlr.press/...atreya25a
24. ^Svitak, T. "The Czech Play That Gave Us the Word 'Robot'." *The MIT Press Reader*, 2019. thereader.mitpress.mit.edu/origin-word-robot-rur
25. ^Devol, G. C., Jr. "Programmed Article Transfer." US Patent 2,988,237, filed December 10, 1954, issued June 13, 1961. patents.google.com/...en
26. ^The Henry Ford. "Robot, First Unimate Robot Ever Installed on an Assembly Line, 1961." thehenryford.org/...183434
27. ^Nilsson, N. J., ed. *Shakey the Robot.* SRI International, 1984, documenting the 1966-1972 project. sri.com/...shakey-the-robot-pub
28. ^Brooks, R. A. "Intelligence without Representation." *Artificial Intelligence* 47, nos. 1-3 (1991): 139-159. courses.csail.mit.edu/...brooks.pdf
29. ^International Federation of Robotics. "World Robotics 2025 report: Global Factory Robot Workforce More Than Doubled in 10 Years." Press release, September 25, 2025. ifr.org/...mand-in-factories-doubles-over-10-years
30. ^International Federation of Robotics. "World Robotics 2025 report: Service Robots See Global Growth Boom." Press release, October 7, 2025. ifr.org/...service-robots-see-global-growth-boom
31. ^International Federation of Robotics. "Humanoid Robots: Vision and Reality." Position paper press release, August 14, 2025. ifr.org/...sion-and-reality-paper-published-by-ifr
32. ^Manufacturing Dive. "US robotics installations rebounded in 2025, on track for more growth: IFR." June 26, 2026. manufacturingdive.com/...823874
33. ^Kim, M. J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., Rafailov, R., et al. "OpenVLA: An Open-Source Vision-Language-Action Model." *Proceedings of the 8th Conference on Robot Learning*, 2024. arxiv.org/...2406.09246
34. ^Physical Intelligence, Black, K., Brown, N., Darpinian, J., Dhabalia, K., Driess, D., Finn, C., Hausman, K., Levine, S., et al. "pi_0.5: a Vision-Language-Action Model with Open-World Generalization." arXiv:2504.16054, April 22, 2025. arxiv.org/...2504.16054
35. ^Physical Intelligence. "pi 0.7: a Steerable Model with Emergent Capabilities." April 16, 2026. pi.website/...pi07
36. ^Gemini Robotics Team, Google DeepMind. "Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer." arXiv:2510.03342, October 2025. arxiv.org/...2510.03342
37. ^Google DeepMind. "Gemini Robotics 2 brings whole body intelligence to robots." July 30, 2026. deepmind.google/...ole-body-intelligence-to-robots
38. ^Khazatsky, A., Pertsch, K., et al. "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset." *Robotics: Science and Systems XX*, 2024. roboticsproceedings.org/...p120.pdf
39. ^Fu, Z., Zhao, T. Z., and Finn, C. "Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation." arXiv:2401.02117, January 2024. mobile-aloha.github.io
40. ^Chi, C., Xu, Z., Pan, C., Cousineau, E., Burchfiel, B., Feng, S., Tedrake, R., and Song, S. "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots." *Robotics: Science and Systems XX*, 2024. arxiv.org/...2402.10329
41. ^Lin, F., Hu, Y., Sheng, P., Wen, C., You, J., and Gao, Y. "Data Scaling Laws in Imitation Learning for Robotic Manipulation." arXiv:2410.18647, October 2024. arxiv.org/...2410.18647
42. ^Liu, B., Zhu, Y., Gao, C., Feng, Y., Liu, Q., Zhu, Y., and Stone, P. "LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning." *Advances in Neural Information Processing Systems 36, Datasets and Benchmarks Track*, 2023. arxiv.org/...2306.03310
43. ^Fei, S., Wang, S., Shi, J., Dai, Z., Cai, J., Qian, P., Ji, L., et al. "LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models." arXiv:2510.13626, October 2025. arxiv.org/...2510.13626
44. ^Li, X., Hsu, K., Gu, J., Pertsch, K., Mees, O., Walke, H. R., Fu, C., et al. "Evaluating Real-World Robot Manipulation Policies in Simulation." *Proceedings of the 8th Conference on Robot Learning*, PMLR 270, 2024. arxiv.org/...2405.05941
45. ^Arkhangelskiy, S. "PhAIL: A Real-Robot VLA Benchmark and Distributional Methodology." arXiv:2605.29710, May 28, 2026. arxiv.org/...2605.29710
46. ^Badithela, A., Snyder, D., Zha, L., Mikhail, J., O'Kelly, M., Dixit, A., and Majumdar, A. "Reliable and Scalable Robot Policy Evaluation with Imperfect Simulators." arXiv:2510.04354, October 2025. arxiv.org/...2510.04354
47. ^Rudin, N., Hoeller, D., Reist, P., and Hutter, M. "Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning." *Proceedings of the 5th Conference on Robot Learning*, PMLR 164 (2022): 91-100. arxiv.org/...2109.11978
48. ^Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., and Hutter, M. "Learning Quadrupedal Locomotion over Challenging Terrain." *Science Robotics* 5, no. 47 (2020). science.org/...scirobotics.abc5986
49. ^Wensing, P. M., Wang, A., Seok, S., Otten, D., Lang, J., and Kim, S. "Proprioceptive Actuator Design in the MIT Cheetah: Impact Mitigation and High-Bandwidth Physical Interaction for Dynamic Legged Robots." *IEEE Transactions on Robotics* 33, no. 3 (2017): 509-522. ieeexplore.ieee.org/...7815333
50. ^Yuan, W., Dong, S., and Adelson, E. H. "GelSight: High-Resolution Robot Tactile Sensors for Estimating Geometry and Force." *Sensors* 17, no. 12 (2017): 2762. mdpi.com/...2762
51. ^Meta AI (FAIR). "Advancing embodied AI through progress in touch perception, dexterity, and human-robot interaction." Blog and Digit 360 release, November 2024. ai.meta.com/...fair-robotics-open-source
52. ^NVIDIA. "NVIDIA Blackwell-Powered Jetson Thor Now Available, Accelerating the Age of General Robotics." Press release, August 25, 2025. nvidianews.nvidia.com/...e-age-of-general-robotics
53. ^Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., et al. "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances." *Proceedings of the 6th Conference on Robot Learning*, 2022. arxiv.org/...2204.01691
54. ^International Organization for Standardization and International Electrotechnical Commission. *ISO/IEC TR 5469:2024 Artificial intelligence - Functional safety and AI systems.* 2024. iso.org/...81283
55. ^Association for Advancing Automation. *ANSI/RIA R15.08-1-2020* and *ANSI/A3 R15.08-2-2023, American National Standard for Industrial Mobile Robots - Safety Requirements.* automate.org/...ems-and-applications-now-available
56. ^Amazon. "Amazon deploys over 1 million robots and launches new AI foundation model." July 2025. aboutamazon.com/...lion-robots-ai-foundation-model
57. ^Intuitive Surgical. "Intuitive Announces Fourth Quarter Earnings." January 22, 2026. isrg.intuitive.com/...es-fourth-quarter-earnings-5
58. ^Deere and Company. "John Deere Reveals New Autonomous Machines and Technology at CES 2025." January 2025. deere.com/...autonomous-9RX
59. ^Lely. "Automatic Milking Robot: Lely Astronaut A5 Next." lely.com/...astronaut
60. ^The Robot Report. "ISO 10218 industrial robot safety standard receives major overhaul." February 18, 2025. therobotreport.com/...dard-receives-major-overhaul
61. ^NVIDIA. "Simplify Generalist Robot Policy Evaluation in Simulation with NVIDIA Isaac Lab-Arena." NVIDIA Technical Blog. developer.nvidia.com/...lab-arena
62. ^Wang, Z., Wang, B., Zhang, H., Du, T., Chen, T., Sun, G., He, Y., Shen, Z., Ye, W., and Li, A. "Vision-Language-Action in Robotics: A Survey of Datasets, Benchmarks, and Data Engines." arXiv:2604.23001, April 24, 2026. arxiv.org/...2604.23001