Sense-Plan-Act Pipeline in Robotics Architecture

The Sense-Plan-Act (SPA) pipeline is one of the foundational control architectures in autonomous robotics, structuring robot behavior into three sequential, discrete phases. Originating in symbolic AI research at MIT and Carnegie Mellon in the 1980s, it remains a reference model against which all subsequent architectures — reactive, hybrid, and behavior-based — are explicitly positioned. Understanding where SPA applies, where it fails, and how it compares to alternative paradigms is essential for engineers specifying robotics architecture systems across industrial, mobile, and autonomous domains.

Definition and scope

The Sense-Plan-Act pipeline defines a serial processing loop in which a robot first acquires environmental data through its sensors, constructs or updates an internal world model, generates a plan based on that model, and then executes motor commands to carry out that plan. Each phase completes before the next begins — a deliberative, top-down approach rooted in classical AI.

The architecture is formally classified within the IEEE Robotics and Automation Society's taxonomy of autonomous system control as a deliberative architecture, distinguishing it from reactive architectures (which bypass the planning phase) and hybrid architectures (which combine both). The Robotics Technology Ontology maintained by NASA's Jet Propulsion Laboratory similarly catalogs SPA as a reference model for knowledge-based autonomy frameworks (NASA JPL Robotics).

Scope-wise, SPA is most applicable when:

It is less suited — and in practice often replaced or augmented — when sensor-to-actuator response times must fall below 10 milliseconds, a threshold relevant to real-time operating systems in robotics.

How it works

The three phases of the SPA pipeline operate as follows:

  1. Sense: Sensor data (LiDAR, cameras, IMUs, force-torque sensors) is acquired and pre-processed. Raw data undergoes filtering, calibration correction, and feature extraction. In modern implementations, this stage incorporates sensor fusion architecture techniques — Kalman filtering, particle filtering, or deep neural network classifiers — to produce a unified environment representation.

  2. Plan: The planning module ingests the world model produced by the sensing phase and applies search, optimization, or symbolic reasoning algorithms to generate an action sequence. Common planners include A*, RRT (Rapidly-exploring Random Trees), and PDDL (Planning Domain Definition Language)-based solvers. PDDL is formalized by the International Planning Competition (IPC), which has defined planning language standards since 1998 (IPC Planning Standards). The output is an explicit plan — a sequence of waypoints, manipulation primitives, or task steps — passed to the execution stage.

  3. Act: The actuator subsystem executes motor commands corresponding to the plan. Closed-loop controllers (PID, model-predictive control) operate here, but at the level of individual actuator commands rather than high-level goal reasoning. This stage maps directly to robot control systems design concerns.

The loop then repeats: after acting, the robot senses again, updating its world model to reflect the consequences of its actions.

Common scenarios

SPA pipeline architecture appears in practice across four primary deployment contexts:

Warehouse and logistics automation: Autonomous mobile robots (AMRs) in facilities operated by large distribution networks often implement SPA variants in which map-based localization (SLAM) feeds a central path planner. The warehouse logistics robotics architecture literature documents planning cycle times of 50–200 milliseconds as typical for structured indoor environments.

Surgical and medical robotics: Deliberative planning is required where pre-operative imaging data informs intraoperative motion sequences. Surgical robotic systems subject to FDA 21 CFR Part 820 (Quality System Regulation) rely on verified, auditable plan generation — a requirement SPA's explicit planning step satisfies architecturally (FDA 21 CFR Part 820).

Space and planetary exploration: NASA's Mars rovers (Spirit, Opportunity, Curiosity, Perseverance) implement deliberative pipelines in which onboard planners generate traversal sequences from terrain models, with round-trip communication delays of 3–22 minutes making real-time Earth-in-the-loop control impossible. The CLARATY (Coupled Layer Architecture for Robotics Autonomy) framework, developed at JPL, extended classical SPA with layered execution monitoring (NASA CLARATY).

Industrial manipulation: Robotic arms performing assembly, welding, or pick-and-place in structured cells use motion planners (e.g., MoveIt in ROS) operating on CAD-derived world models. The motion planning architecture for such systems is a direct implementation of the Plan phase.

Decision boundaries

Selecting SPA over alternative architectures requires evaluating four structural decision factors:

Latency tolerance: SPA introduces cumulative latency across all three phases. If environmental conditions change faster than a complete SPA cycle, the plan becomes stale before execution completes. Reactive architectures, as described in reactive vs deliberative architecture, address this by eliminating the planning phase entirely for time-critical reflexes.

World model fidelity: SPA depends on the sensing phase producing an accurate, current world model. In dynamic environments with moving obstacles or unpredictable humans, model staleness causes plan failures. Hybrid architecture in robotics compensates by pairing a deliberative planner for high-level goals with a reactive layer for immediate collision avoidance.

Computational resources: Full deliberative planning — particularly probabilistic roadmap or sampling-based planners — can consume 100% of a CPU core for hundreds of milliseconds per planning cycle. Embedded platforms with constrained processors, covered in embedded systems robotics architecture, may not support this load.

Auditability and safety certification: For systems subject to ISO 10218-1 (industrial robot safety) or IEC 62061 (functional safety), explicit plan generation with logged decision states satisfies audit requirements more readily than emergent reactive behavior (ISO 10218-1 overview via ISO.org). This is a decisive advantage in regulated sectors where safety architecture in robotics mandates traceable decision chains.

The SPA pipeline does not inherently support parallel sensing and planning — both phases are serialized by design. Behavior-based robotics architecture, introduced by Rodney Brooks at MIT in 1986, was explicitly developed as a rejection of this serialization in favor of concurrent, layered subsumption.

References

Read Next

Real-Time Operating Systems in Robotics Architecture This page covers the technical definition and scope of RTOS within robotics, the scheduling and interrupt mechanics that... Sensor Fusion Architecture in Robotic Systems The structural decisions made in a fusion architecture directly determine perception latency, fault tolerance, and the... Control Systems Design in Robotics Architecture The discipline spans classical control theory, modern state-space methods, and AI-augmented approaches, each carrying distinct...