> For the complete documentation index, see [llms.txt](https://neurosymbolicai.gitbook.io/neuro-symbolic-ai-in-practice/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://neurosymbolicai.gitbook.io/neuro-symbolic-ai-in-practice/part-ii-background/chapter-2/2-1-classical-planning.md). # 2.1 Classical Planning Milestones AI Planning is not a speculative research direction. It is a mature engineering discipline with a track record of real deployments in some of the most demanding domains imaginable: spacecraft operations, planetary exploration, and mission-critical satellite scheduling. The systems described in this section operated in environments where failure was not an option — and they succeeded. ## Remote Agent — NASA Deep Space One (1998–1999) The most dramatic early milestone in autonomous AI Planning is the **Remote Agent** experiment aboard NASA's Deep Space One spacecraft.(Muscettola et al., 1998) In May 1999, for the first time in history, a computational system was given **full autonomous control of a spacecraft in flight**. Remote Agent — a fully autonomous planning and scheduling system developed jointly by NASA Ames Research Center and Jet Propulsion Laboratory — commanded DS1's attitude control, power management, fault detection, and scientific instrument scheduling without ground-based intervention for two extended sessions totaling approximately 21 hours. What made Remote Agent remarkable was not just its autonomy, but its architecture. The system integrated three tightly coupled symbolic components: * A **planner** (built on top of HSTS — Heuristic State and Transition Scheduling)(Muscettola, 1993) that generated and maintained a plan of activities over a rolling time horizon. * An **executive** that monitored plan execution and detected deviations from expectations. * A **fault recovery system** that triggered replanning when unexpected events occurred — including, during the experiment, deliberate fault injection by ground controllers to stress-test the system. Remote Agent successfully handled induced fault scenarios, replanned in real time, and restored nominal operations. It demonstrated that formal planning systems could operate reliably in a radiation-hardened, communication-latency-challenged, genuinely adversarial environment.(Muscettola et al., 1998) > **Significance:** Remote Agent demonstrated that autonomous planning is not merely theoretically sound — it is operationally reliable in the hardest environments humans have ever built systems for. *References:*\ Muscettola, Nicola, P. Pandurang Nayak, Barney Pell, and Brian C. Williams. "Remote Agent: To Boldly Go Where No AI System Has Gone Before." *Artificial Intelligence* 103.1–2 (1998): 5–47. *** ## Hubble Space Telescope — SPIKE (1993–present) The **SPIKE** scheduling system has been managing observations for the Hubble Space Telescope since 1993 — one of the longest-running AI planning deployments in history.(Johnston & Miller, 1994) The Hubble scheduling problem is genuinely hard. The telescope must schedule **10,000+ scientific observations per year** from a queue of competing proposals, while respecting telescope orientation constraints (the Sun and Moon must remain in protected zones), managing thermal constraints, handling guide star availability and gyroscope limitations, and satisfying fairness constraints among competing science programs from the global astronomy community. Before SPIKE, human schedulers required **weeks** to construct an observation schedule for a single orbit visibility window. SPIKE reduced this to **hours**, and ultimately to near-real-time operation. The system uses a constraint-based scheduling approach built on top of classical planning primitives, with specialized heuristics for astronomical domain constraints.(Johnston & Miller, 1994) SPIKE has been continuously operational for over three decades — longer than any other AI planning system in operational deployment. It has scheduled observations that produced some of the most iconic scientific results in the history of astronomy: the Hubble Deep Field images, measurements confirming the accelerating expansion of the universe, and detailed atmospheric characterization of exoplanets. > **Significance:** SPIKE proved that AI Planning systems are maintainable, evolvable, and reliable at production scale over multi-decade time horizons. *Reference:* Johnston, Mark D., and Glenn Miller. "Spike: Intelligent Scheduling of Hubble Space Telescope Observations." *Innovative Applications of Artificial Intelligence (IAAI)*, 1994, pp. 1–8. *** ## Mars Exploration Rovers — MAPGEN (2003–2010) The twin Mars Exploration Rovers Spirit and Opportunity — designed for 90-day missions, operational for years (Spirit until 2011 (last contact March 2010; mission declared over May 2011), Opportunity until 2018 (last contact June 2018; declared lost February 2019)) — were managed in part by **MAPGEN**: the Mixed-Initiative Activity Planning Generator.(Ai-Chang, 2004) The Mars rover planning problem combines several challenges that classical planning handles elegantly but that would be extremely difficult to address with a purely learned system. Each Martian day (sol), the rover operates under a fixed power budget from its solar panels, a fixed data volume budget, and a thermal operating window — three resource constraints that must be satisfied simultaneously. The rover's position after a drive is not exactly known, requiring contingency planning. Science team members from multiple institutions must negotiate competing priorities within a single 24-hour planning cycle. MAPGEN generated daily activity sequences for the rovers, incorporating planner output with human scientist oversight in a mixed-initiative interface. The planner ensured resource constraint satisfaction; human scientists provided goal priorities and terrain judgment. The combination — human knowledge plus formal guarantees — is a direct precursor of the LLM-Modulo framework described in [§1.3](/neuro-symbolic-ai-in-practice/part-i-motivation/chapter-1/1-3-llm-modulo.md).(Bresina et al., 2005) > **Significance:** MAPGEN demonstrated that AI Planning scales to multi-year real-world deployments under genuine resource constraints, and that human–planner collaboration is a productive design pattern. *References:*\ Ai-Chang, Michael, et al. "MAPGEN: Mixed-Initiative Planning and Scheduling for the Mars Exploration Rover Mission." *IEEE Intelligent Systems* 19.1 (2004): 8–12. \ Bresina, John L., et al. "Activity Planning for the Mars Exploration Rovers." *Proceedings of ICAPS*, 2005. *** ## CASPER — Earth Observing One (2000–2017) The **CASPER** (Continuous Activity Scheduling Planning Execution and Replanning) system managed scheduling for NASA's Earth Observing One (EO-1) satellite — a technology demonstration mission that became the first fully autonomous satellite scheduler.(Chien et al., 2000) EO-1 carried the Hyperion hyperspectral imager and the Advanced Land Imager, generating massive volumes of data that had to be carefully scheduled against downlink opportunities, power constraints, and onboard storage limits. CASPER provided **incremental, continuous replanning** — unlike traditional one-shot schedulers, CASPER maintained a rolling schedule that updated in near real-time as new opportunities and constraints emerged. The key innovation was *reactive scheduling*: CASPER detected schedule violations as they occurred (due to cloud cover, unexpected power consumption, or new science priorities) and repaired the schedule locally rather than recomputing from scratch. This is a form of *plan repair* — a technique now central to autonomous agent design.(Chien et al., 2000) > **Significance:** CASPER demonstrated that AI Planning can operate in a reactive, continuous mode — not just as a one-shot batch process — enabling autonomous adaptation to dynamic real-world environments. *Reference:* Chien, Steve, et al. "Using Iterative Repair to Increase the Responsiveness of Planning and Scheduling for Observation Scheduling." *Journal of Autonomous Agents and Multi-Agent Systems* 4.1–2 (2000): 129–145. > **Next:** [§2.2 — Modern Neuro-Symbolic Systems](/neuro-symbolic-ai-in-practice/part-ii-background/chapter-2/2-2-neuro-symbolic-systems.md) moves from purely classical deployments to hybrid systems that combine neural components with formal planning — from LLM+P through AlphaProof and HAIMEDA. --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter: ``` GET https://neurosymbolicai.gitbook.io/neuro-symbolic-ai-in-practice/part-ii-background/chapter-2/2-1-classical-planning.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.