Loading field map
Building a robot that can act intelligently in the physical world forces a fundamental choice: how much should the robot rely on an internal model of the world, and how much on direct interaction with its environment? This tension between deliberation and reactivity has driven the history of AI robotics since the 1960s, producing a series of frameworks that each prioritized different answers.
The earliest influential framework, Deliberative Robotics (1966–1990), assumed that intelligent action requires an explicit symbolic world model and a plan. The archetype was Shakey the Robot at SRI, which used STRIPS (a theorem-proving approach) to generate plans by searching a logical representation of the world. The robot would sense the environment, update its model, plan a sequence of actions, and then execute them — this sense-plan-act cycle became the standard template. Deliberative robotics placed its faith in the power of internal representation and logical reasoning. Its central limitation, however, was brittleness: the world changes faster than the robot can replan, and building a complete, accurate model proved infeasible in all but the most controlled settings.
Behavior-Based Robotics (1986–2005) emerged as a direct challenge to the deliberative approach. Rodney Brooks's subsumption architecture rejected central models and planning altogether, arguing that intelligence must be embodied and situated. Instead of a single reasoning engine, the robot was controlled by a layered set of simple, reactive behaviors — such as "avoid obstacles" or "follow light" — that competed and cooperated without any internal representation. This framework prioritized real-time response and robustness over optimality. Behavior-based robotics showed that much can be achieved without explicit deliberation, but it also struggled with tasks requiring long-term planning or reasoning about unobserved parts of the environment.
Robot Learning (1993–Present) offered a third path: instead of hand-coding either a world model or a set of reactive rules, the robot could acquire its own control policies through interaction. Early work applied reinforcement learning (RL) to tasks like grasping and navigation, often using trial and error to learn state-action mappings. This framework shifted the emphasis from engineering to learning, but initially suffered from sample inefficiency and lacked the ability to handle high-dimensional state spaces. Robot Learning coexisted with behavior-based methods, sometimes absorbing their reactive insights into learned policies, and later provided the foundation for Deep Reinforcement Learning, which became a dominant force.
Probabilistic Robotics (1997–Present) transformed the field by providing a rigorous mathematical foundation for handling uncertainty — something that both deliberative and reactive frameworks had largely overlooked. Instead of assuming perfect sensors and deterministic actions, probabilistic robotics uses Bayesian filtering to estimate the robot's state from noisy data. Landmark achievements such as simultaneous localization and mapping (SLAM) became tractable. This framework did not replace earlier approaches; rather, it supplied an infrastructure on which both deliberative planners and learned controllers could be built. Today, probabilistic methods are a standard tool in localization, mapping, and sensor fusion.
Two frameworks expanded the design space in orthogonal directions. Evolutionary Robotics (1994–Present) uses population-based search (genetic algorithms) to evolve robot controllers and morphologies offline, without requiring the robot to learn during its lifetime. This contrasts with Robot Learning's online, individual acquisition, and is especially useful when the learning objective is unclear or when the robot's physical form itself is subject to optimization. Swarm Robotics (1995–Present) draws inspiration from insect colonies: simple robots with local interactions produce global collective behaviors. It extends the behavior-based principle of decentralization to multi-robot systems, emphasizing scalability and robustness over individual intelligence.
Cognitive Robotics (1999–Present) marked a partial revival of internal representation, but now informed by the critiques from behavior-based robotics. Rather than assuming a complete, static world model, cognitive robotics integrates high-level reasoning with perception and action, often using hybrid architectures that combine reactive layers with deliberative planning. It also absorbs ideas from probabilistic robotics to handle uncertainty in reasoning. Cognitive robotics addresses tasks that require explicit awareness of goals, beliefs, and intentions, such as human-robot interaction or complex manipulation.
Developmental Robotics (2001–Present) takes inspiration from human cognitive development: the robot starts with simple innate behaviors and gradually acquires more complex skills through staged, open-ended learning. This framework synthesizes ideas from Robot Learning and Cognitive Robotics, but with a focus on the process of development itself — how a robot can discover its own body, explore its environment, and build representations over time. Developmental robotics remains a live research program, pushing toward lifelong learning and autonomous skill acquisition.
Today, no single framework dominates. Robot Learning — particularly deep reinforcement learning and imitation learning — is at the forefront of many applications, from autonomous driving to dexterous manipulation. Probabilistic Robotics is an accepted foundation for state estimation in nearly all robotic systems. Cognitive Robotics and Developmental Robotics are active but more specialized, while Evolutionary Robotics and Swarm Robotics serve niches where population-based design or collective behavior are essential. Behavior-based principles live on in many low-level control modules. The leading frameworks largely agree that handling uncertainty is crucial, that learning from data is powerful, and that hybrid architectures often outperform pure approaches. They disagree, however, on the necessity of explicit representation: some argue that end-to-end learning can internalize any needed structure, while others maintain that reasoning with symbols and models remains indispensable for tasks requiring logical coherence or safety guarantees. The tension first seen between deliberative and behavior-based robotics has not been resolved so much as transformed into a productive division of labor, with each framework contributing its strongest tools to a common engineering practice.