Plan your Self-Adaptation! Ontology-Mediated Task and Architecture Co-adaptation Planning for Robots

Gustavo Rezende Silva, Tobias John, Francisco Martín, EinarR Broch JohnsenarXiv 2026 · arXiv

TL;DR

PLANTA is a new framework that allows robots to jointly plan their high-level tasks and low-level architectural configurations (like sensor modes or algorithms) in one go. By automatically translating architectural knowledge from ontologies into planning logic, it enables robots to adapt to failures and environmental changes 3x faster and with significantly less memory than previous state-of-the-art methods.

What Problem Does This Paper Address?

Autonomous robots operating in the real world—like underwater drones or warehouse AGVs—face a constant need to adapt. This adaptation happens on two interdependent levels:

  1. Task Planning: Deciding what actions to take to reach a goal.
  2. Architecture Reconfiguration: Deciding how to set up the robot's internal systems (which sensors to use, which localization algorithm to run) to perform those actions.

The "gap" is that these two are often handled separately. A task planner might pick a path without knowing if the current sensor setup can handle the lighting in a specific corridor. Conversely, reactive systems might only reconfigure once a failure occurs, which is often too late. Existing "co-adaptation" approaches that try to solve both simultaneously often struggle with "state-space explosion," becoming too slow or memory-intensive for real-time use on actual hardware.

Methods

The authors introduce PLANTA, an ontology-mediated framework that treats task and architecture co-adaptation (TACA) as a single planning problem.

Key components include:

  • PDDL-TOMASys Ontology: A model that describes robot functionalities, the "designs" (architectural variants) that realize them, and the quality requirements (e.g., "needs high visibility") for each.
  • Owl2PDDL Tool: A custom translator that converts Web Ontology Language (OWL) axioms into PDDL derived predicates. This allows the planner to "reason" about the robot's internal state using standard planning languages.
  • Managed Actions: The system automatically wraps standard task actions (like move) into "managed" versions. These managed actions include preconditions that ensure the robot is in a valid, high-utility configuration before the action starts.
  • Satisficing Planning: Instead of searching for a mathematically perfect "optimal policy" for every possible future (which is computationally expensive), PLANTA uses satisficing planning. It finds a "good enough" plan quickly and simply replans if the environment changes or a precondition is violated.

The entire workflow is integrated into the ROS 2 ecosystem, specifically extending the PlanSys2 framework to support complex PDDL features like existential preconditions and derived predicates.

Fig. 1. PLANTA framework, organized into design time, deployment time, and runtime, separated by dashed vertical lines. Some PlanSys2 nodes are omitted for clarity, and the Planner and Plan Execution components shown were extended as part of this work.

Datasets

The framework was evaluated using two primary robotic simulation scenarios:

  1. SUAVE (Autonomous Underwater Vehicle): A pipeline inspection mission where the robot faces thruster failures, changing water visibility, and battery depletion.
  2. AGV Navigation: A ground robot navigating a grid-like environment with varying corridor types (dark, unsafe, dusty) that require different sensor configurations.

Key Insights and Ideas

  • Ontology as a Bridge: Using ontologies for architectural modeling allows engineers to define system capabilities in a human-readable, reusable way, while the automated translation to PDDL ensures the robot can actually use that knowledge for planning without manual coding.
  • The Power of "Good Enough": A major "aha moment" is the trade-off between optimality and tractability. By favoring satisficing plans over optimal ones, PLANTA remains functional in complex scenarios where optimal planners simply run out of memory or take minutes to find a solution.
  • Local vs. Global Utility: PLANTA simplifies the math by selecting the best configuration for a specific function (local utility) rather than trying to optimize the entire mission's energy or time (global utility). While technically sub-optimal, the results show the resulting plans are nearly identical in quality to optimal ones in practical scenarios.

Key Results

  • Performance Efficiency: In the AGV navigation tests, PLANTA was 3.00 times faster at the median and used 6.84 times less peak memory than the leading optimal-policy approach.
  • Scalability: As the number of mission-level actions and map size increased, PLANTA's planning time remained manageable (under 15 seconds for complex stress tests), while optimal planning became impractical for online use.
  • Plan Quality: Despite not seeking mathematical optimality, PLANTA produced plans with near-identical path lengths and battery consumption compared to optimal methods.
  • Robustness: In the underwater SUAVE scenario, PLANTA successfully handled concurrent thruster failures and visibility changes, outperforming simpler reactive or random adaptation strategies.

Limitations and Future Work

  • Local Optimization Limits: Because PLANTA optimizes configurations locally per function, it might miss "global" wins—like choosing a slightly less efficient configuration now to save a specific sensor's life for a critical task later.
  • Reconfiguration Costs: The current model doesn't explicitly account for the time or energy "cost" of the act of switching configurations (e.g., the 5 seconds it takes for a sensor to boot up).
  • Numeric Reasoning: The reliance on derived predicates makes it difficult to handle continuous numeric values (like exact battery voltage) without symbolic encoding.

Future work aims to incorporate action costs to model reconfiguration overhead and explore ways to support mission-level objective optimization without sacrificing the speed of the current approach.

Summary updated

Abstract

Autonomous robots may need to adapt both their task plans and architectural configurations as operating conditions change. These decisions are interdependent: architectural configurations determine which actions a robot can perform, while task plans determine which capabilities are required. This paper presents PLANTA, an ontology-mediated approach to task and architecture co-adaptation that considers these dependencies during mission planning. PLANTA uses an ontology expressed in the Web Ontology Language to represent architectural alternatives and their associated functional and quality requirements, while the Planning Domain Definition Language defines task actions and mission goals. By automatically translating and integrating these models, PLANTA enables joint planning of task and reconfiguration actions. The approach uses satisficing planning and replans when changes during execution invalidate the current plan. PLANTA is evaluated in two robotic scenarios involving underwater inspection and ground navigation. The results support the feasibility of runtime co-adaptation and show lower planning time and memory consumption compared with optimal-policy generation in the evaluated navigation scenario, while producing plans of similar quality.
BibTeX
@misc{silva2026plan,
  author = {Gustavo Rezende Silva and Tobias John and Francisco Martín and EinarR Broch Johnsen},
  title = {{Plan your Self-Adaptation! Ontology-Mediated Task and Architecture Co-adaptation Planning for Robots}},
  year = {2026},
  eprint = {2611.25254},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.25254},
}
Figure