Robotics

Summary

Robotics is the engineering discipline concerned with machines that sense their surroundings, decide what to do, and act on the physical world. It covers mechanical design, actuation, sensing, control, motion planning, safety, and increasingly the trained behavior that replaces what an engineer would once have written by hand. Two economies run inside the field at once: millions of installed machines performing motion that was specified in advance, and a frontier where robots are expected to handle variation nobody specified. The dividing line is not the hardware but whether the required behavior can be written down or has to be recovered from recorded physical experience.

Robotics: Definition and Scope

Robotics is the design, construction, and operation of programmable machines that act physically. The International Federation of Robotics publishes the ISO 8373 vocabulary definition of a robot as “a programmed actuated mechanism with a degree of autonomy to perform locomotion, manipulation or positioning”. The same vocabulary defines an industrial robot as an “automatically controlled, reprogrammable, multipurpose manipulator, programmable in three or more axes, which can be either fixed in place or fixed to a mobile platform for use in automation applications in an industrial environment”. Those two definitions carry the field’s central distinction: every robot is actuated and programmed, and what varies is how much autonomy the decision layer holds.

The discipline covers manipulators, mobile robots including autonomous mobile robots in warehouses, legged and humanoid platforms, aerial and underwater vehicles, surgical and rehabilitation systems, and self-driving vehicles as a specialized branch. It also covers the layers that surround the machine: kinematics and dynamics, state estimation, motion planning, real-time control, calibration, human robot interaction, and functional safety.

Robotics is explicitly not a synonym for artificial intelligence applied to machines. It predates modern machine learning by decades and most deployed robots today contain little or none of it. Software that produces only digital outputs is out of scope, as is fixed automation with no programmable decision layer.

The industry term and the academic term diverge in emphasis. In industry, “robotics” often means the automation stack around an arm or a fleet: integration, tooling, cell design, safety, uptime. In academia it more often means the theory underneath: control, estimation, planning, and learning. Physical AI is the current direction inside the discipline in which learned components take over functions that were previously engineered, and Embodied AI is the academic framing of intelligence arising through a body acting in an environment. Robotics is the parent discipline that both sit inside.

How the Field Arrived Here

The pattern the table reveals is that each step forward came from removing a dependency on a human specification: first the fixed cam and linkage, then the teach pendant path, then the hand engineered perception pipeline, and now the hand engineered manipulation skill. Every removal moved the burden onto recorded evidence about how the physical world actually behaved.

Why Now

Robotics has been a working industry since 1961. What changed recently is which parts of a robot’s behavior can be obtained rather than authored.

Hardware stopped being the gate on experimentation. Unitree announced the G1 humanoid at $16,000 in May 2024, a research capable legged platform at roughly the price of a used car. Cheap hardware does not make a robot useful, but it changes how many teams can generate physical experience at all, and experience is the input the learned layer consumes.

A shared software substrate matured. The middleware, simulators, dataset formats, and tooling that took a decade to accumulate after ROS mean a new robotics team no longer writes its own transport, coordinate frame handling, or logging. That reuse is what makes cross-team data exchange conceivable in the first place.

Learned components reached the manipulation layer. Vision-Language-Action Models gave many groups one formulation for conditioning actions on observations and instructions, so the skill that was previously scripted per task can be trained instead. This is a continuation of the substrate point rather than a separate factor: shared tooling and shared datasets are what made the training runs reproducible across labs.

Capital arrived at scale. Crunchbase reported on 22 June 2026 that robotics startups had raised $18.8 billion during 2026 to that date, against $15 billion across all of 2025. Individual rounds reflect the same pattern, with Agility Robotics announcing a $2.5 billion pre-money valuation and over $620 million in gross transaction proceeds on 24 June 2026. These figures cover robotics startups broadly and say nothing about whether any individual program is adequately funded.

Safety and standards caught up enough to permit deployment. The third edition of ISO 10218-1 was published on 5 February 2025, replacing a 2011 standard written before collaborative operation and mobile manipulation were common. Deployment in a regulated factory depends on there being a current standard to certify against.

None of this solved the part that matters most for the frontier. Cheaper hardware, shared middleware, trained policies, capital, and updated standards do not by themselves produce trustworthy recorded experience. They do not fix inconsistent calibration across cells, incompatible action definitions between platforms, missing outcome labels, or the gap between an offline score and what the robot does on the tenth attempt.

Robotics Compared to Physical AI

The two terms are frequently used interchangeably, which obscures a distinction practitioners depend on.

The row that causes the others is where behavior comes from. Once behavior is fit to data rather than written, the artifact you maintain changes, the way you establish correctness changes, and the thing you inspect when the robot misbehaves changes with it. This is a difference in degree rather than a clean split. Nearly every deployed system in 2026 is a hybrid, with a learned policy running inside a classical stack that still owns state estimation, motion limits, and safety.

Why Robotics Matters

The frontier problem in robotics is not intelligence in the abstract. It is that a robot’s competence is bounded by physical experience that somebody has to produce, and that experience is expensive and quietly fragile.

Consider the task BMW Group put Figure 03 to at Plant Spartanburg: picking unsorted components out of larger containers and sorting them into sequencing trolleys so parts arrive at assembly in sequence. To learn that behavior rather than script it, a training set has to carry, for every attempt: images from each mounted and onboard camera, depth, the full joint state, the end effector pose expressed in a known frame, gripper state and contact force, the instruction or target part identity, a shared time base across all of it, the calibration that relates each camera to the robot base at the moment of capture, and whether the placement succeeded.

Here is where it breaks, and it is specific to physical machines rather than to models. A camera on the cell is bumped and recalibrated partway through a collection campaign. Nobody records the change, because the calibration lives in a config file on the cell controller rather than in the dataset. Every trajectory after that point expresses end effector poses in a base frame that no longer matches the images. The files are complete, the timestamps are clean, the training loss falls normally, and frame by frame review shows nothing wrong, because each individual frame is correct. The policy learns a systematic spatial offset that is real in half the data and absent in the other half. On hardware it reaches slightly past the part, grips air or the container wall, and the failure looks like a grasping weakness rather than a data defect.

The same class of defect appears at the schema level. One cell records actions as joint velocity commands, another as end effector delta poses, a third as absolute target poses; one treats the gripper as binary, another as continuous width. Pooled into a single training set without normalization, these teach the model that identical observations warrant contradictory actions. Neither failure is visible in an offline metric, because the held out split carries the same corruption as the training split. Both are only detectable if calibration, coordinate frame conventions, action semantics, and outcome labels are treated as first class dataset content that gets validated before a training run, rather than as environment state that happened to be true on the day.

Industry Perspective

Leading organizations disagree sharply about almost everything visible. NVIDIA builds an open model and simulation stack it expects others to deploy on. Figure AI builds a vertically integrated humanoid and runs it inside a named customer’s plant. Agility Robotics has taken a bipedal logistics robot to more than 65,000 hours of operation across nine customer facilities and over $300 million in multi-year contracted orders before going public. Universal Robots and the established industrial vendors continue to sell arms whose behavior is programmed rather than learned, into a market that installed 542,000 industrial robots in 2024. These are incompatible bets on form factor, business model, and how much autonomy a customer will accept.

They converge on one thing: hours of real operation in a real environment are the asset, and everything is organized to accumulate them.

1. BMW Group and Figure AI ran Figure 02 in the body shop for ten months across production of more than 30,000 BMW X3 vehicles before moving to a new application with the next generation robot.

2. Agility Robotics leads its public market case with cumulative operating hours and facility count rather than with model capability.

3. NVIDIA distributes simulation and world model tooling whose stated purpose is generating and validating physical experience that would otherwise have to be collected.

4. The IFR reported on 7 October 2025 that Robot-as-a-Service fleets grew 31 percent in 2024, a commercial model that keeps the vendor attached to the deployed fleet and its operating record instead of ending the relationship at the sale.

The inference this supports is narrow and worth stating precisely. It does not show that data is the only binding constraint, since hardware reliability, unit economics, control frequency, safety certification, and service logistics all remain material. It shows that the organizations closest to deployment treat the accumulated, labeled record of physical operation as the durable asset, and that this holds across companies that agree on nothing else about how a robot should be built.

Kotwel’s Role

Kotwel does not build robots, design control stacks, or take positions on form factor. That boundary is deliberate. Whether the system is a fenced industrial arm, a mobile manipulator, or a humanoid, learned behavior depends on the quality of the physical experience captured from real operation.

Kotwel turns raw records of physical operation into structured, reliable, and traceable datasets for robot training, evaluation, and continuous improvement. The data operation preserves calibration and coordinate-frame provenance, aligns video, depth, state, force, and action streams on a shared time base, standardizes action representations across cells and platforms, defines consistent annotations and outcome labels, and applies quality assurance at the trajectory level.

The work becomes more important as a robotics program scales. Data collected across different robots, operators, environments, and generations of hardware has to remain interpretable and compatible over time. Kotwel provides the data operations needed to maintain that continuity through consistent schemas, versioning, provenance, and quality controls, so accumulated physical experience can remain useful as datasets expand and models evolve.

Kotwel’s role sits between physical operation and model development: transforming what robots experience in the real world into data that teams can trust, reuse, and build on.

Key Concepts

Physical AI is the direction within robotics in which learned models take over perception, reasoning, or action generation from engineered components.

Embodied AI is the academic framing of intelligence that develops through a body perceiving and acting in an environment.

Autonomous Systems operate with limited direct intervention in dynamic environments, spanning vehicles, aerial platforms, and industrial equipment as well as robots.

Humanoid Robots are general purpose bipedal platforms whose appeal is operating in environments built for people without modifying those environments.

ROS 2 is the middleware and tooling ecosystem most robotics teams build on, providing transport, coordinate frame handling, and logging.

Robot Demonstration Data records tasks performed through teleoperation or other teaching methods in a robot’s own observation and action space.

Vision-Language-Action Models condition robot actions on visual observations and language instructions in a single trained model.

Behavior Cloning is the supervised imitation method underneath most learned manipulation policies, mapping observations to actions from demonstrations.

Sim to Real Transfer studies how capability learned in simulation carries to physical hardware and where the reality gap causes failure.

Synthetic Data for Robotics uses simulated or generated experience to extend scenario coverage beyond what physical collection can reach.

Closed-Loop Policy Evaluation tests a policy through repeated perception, action, and feedback on hardware, exposing compounding errors that offline metrics miss.

Multimodal Dataset Synchronization aligns sensor, state, action, and annotation streams on a shared time base.

Robot Calibration and Coordinate Frames defines the spatial relationships between sensors, links, and tools, and is the provenance most often lost between the cell and the dataset.

Action Space Standardization reconciles the different action representations and gripper semantics used across robot platforms so pooled data does not teach contradictions.

Further Reading

Primary sources

1. World Robotics 2025, industrial robots press release, International Federation of Robotics, 25 September 2025.

2. World Robotics 2025, service robots press release, International Federation of Robotics, 7 October 2025.

3. ISO 10218-1:2025, Robotics, Safety requirements, Part 1: Industrial robots, published 5 February 2025.

4. ISO 10218-2:2025, Robotics, Safety requirements, Part 2: Industrial robot applications and robot cells, published 2025.

5. Robot definitions at ISO, International Federation of Robotics, carrying the ISO 8373 vocabulary definitions.

6. BMW Group advances the use of physical AI in production with the Figure 03 project in Spartanburg, BMW Group, 25 June 2026.

7. Agility Robotics to Go Public Through Merger with Churchill Capital Corp XI, Agility Robotics, 24 June 2026.

8. Open X-Embodiment: Robotic Learning Datasets and RT-X Models, 2023.

Industry and strategy

1. Wizards of ROS: Willow Garage and the Making of the Robot Operating System, IEEE Spectrum, 2019.

2. How KAIST’s DRC-Hubo Won the DARPA Robotics Challenge, IEEE Spectrum, 2015.

3. Unitree Robotics unveils G1 humanoid for $16K, The Robot Report, 15 May 2024.

4. Robotics startup venture funding surges in 2026, Crunchbase News, 22 June 2026.

5. Universal Robots company history, covering the 2008 UR5 launch.