Robotics is the branch of engineering and computer science concerned with the design, construction, operation, and programming of physical machines, robots, that sense their environment and take physical action within it. While robotics predates modern artificial intelligence by decades, tracing back to industrial arms deployed on factory floors starting in the 1960s, the field has become increasingly intertwined with AI as perception, planning, and control have shifted from hand-coded rules toward learned models, a convergence often discussed under the broader heading of Embodied AI.
From industrial to learned robotics
Early industrial robots, such as the Unimate arm installed at a General Motors plant in 1961, operated through fixed, pre-programmed motion sequences in tightly controlled environments and required no perception of an unstructured world. Research robotics through the 1980s and 1990s, exemplified by projects like Rodney Brooks's subsumption-architecture robots, explored reactive, sensor-driven control as an alternative to symbolic planning. The 2004 and 2005 DARPA Grand Challenge competitions, which tasked autonomous vehicles with navigating the desert, are widely credited with jump-starting modern robotic perception and planning research and with seeding the Self-driving car industry. Over the following two decades, Computer vision and Reinforcement learning increasingly replaced hand-engineered rules for tasks like grasping, locomotion, and navigation, allowing robots trained largely in simulation to generalize to some degree in the physical world, a transfer commonly called sim-to-real.
AI-driven perception and control
Modern robotics draws heavily on the same Deep learning techniques used elsewhere in AI. Convolutional neural network and, later, Vision-language model systems give robots the ability to recognize objects and scenes; reinforcement learning and imitation learning, in which a robot learns by mimicking recorded human demonstrations, are used to train control policies; and, since the early 2020s, vision-language-action (VLA) models have combined a large pretrained vision-language-model backbone with an output head that directly produces robot motor commands, letting a single model follow natural-language instructions across varied physical tasks. This has produced a wave of general-purpose robotic foundation models pursued by both established robotics labs and AI companies newly entering the field.
Application domains
Industrial and warehouse robotics remains the largest deployed base, with robotic arms and autonomous mobile robots widely used in manufacturing and logistics, including at large fulfillment operations run by companies such as Amazon. Surgical robotics, exemplified by the da Vinci system, extends a surgeon's precision through teleoperated instruments. Agricultural robots handle harvesting and crop monitoring, and household robots, most successfully robotic vacuum cleaners, represent the most widely adopted consumer category to date. Humanoid robot platforms, aimed at general-purpose tasks in human-designed environments, and Self-driving car systems represent the most capital-intensive and closely watched frontiers of the field as of the mid-2020s.
Challenges
Despite rapid AI progress, robotics has historically advanced more slowly than purely digital AI applications, a gap sometimes summarized as Moravec's paradox: tasks that are trivial for humans, such as picking up an irregularly shaped object or walking over uneven terrain, remain difficult for machines, while tasks that are hard for humans, such as complex calculation, are comparatively easy for computers. Physical robots also face constraints that purely software-based AI systems do not, including hardware cost, durability, safety around humans, and the difficulty of collecting large, diverse real-world training data comparable to the internet-scale text and image corpora that trained modern Large language model and Text-to-image generation systems.