Introduction
In traditional AI robotics, a significant "representation gap" exists between planning and acting. Planning typically relies on descriptive models (e.g., PDDL or STRIPS) that define what an action does via preconditions and effects. Conversely, real-world acting requires operational models—complex procedures that describe how to perform tasks, incorporating loops, conditional logic, and reactive error handling.
This dual-model approach often leads to consistency issues. If a planner assumes an action is feasible but the actor's implementation handles it differently, the system may fail. Furthermore, descriptive models struggle to represent the rich control structures necessary for dynamic, closed-loop decision-making. This paper introduces RAEplan, a system that unifies acting and planning by using a single operational model for both.
The Refinement Acting Engine (RAE) and RAEplan
The proposed integrated system is based on the Refinement Acting Engine (RAE). Instead of abstract logic, RAE uses hierarchical task-oriented models—essentially programs defining multiple "refinement methods" to achieve a task.
Monte Carlo Rollouts
When the actor must choose between different methods, it invokes RAEplan. Unlike traditional state-space search, RAEplan performs "rollouts" by simulating the execution of its own operational procedures. By sampling non-deterministic outcomes of commands, it estimates which method is most likely to succeed efficiently.
Efficiency Metric
Traditional planning often minimizes cost, but in environments with potential failures (infinite cost), this approach falters. The authors introduce an Efficiency ($\nu$) metric, defined as $1/cost$. This allows the planner to maximize a value between 0 and 1, gracefully handling paths that might lead to dead ends.
Evaluation Domains
The authors evaluated RAEplan across four simulated benchmark domains:
- Explorable Environment (EE): Multi-robot coordination in terrain with dead ends.
- Chargeable Robot (CR): Resource management and sensing under uncertainty.
- Spring Door (SD): Collaborative task requiring robots to hold doors open.
- Industrial Plant (IP): A complex workshop with dynamic machine failures and concurrent tasks.
Key Results and Insights
The core "aha moment" of this research is that the actor's code is the planning model. There is no translation layer where information can be lost. By treating planning as a simulation of acting logic, the robot can reason about complex programming constructs like loops and reactive events.
Experimental results across all domains showed that increasing planning effort—controlled by parameters $b$ (breadth) and $k$ (samples)—consistently led to higher average efficiency. Notably, the "retry ratio" dropped sharply as planning was introduced. In the Spring Door domain, for instance, look-ahead helped robots avoid methods likely to fail given the current environment state, significantly saving time and energy.
Conclusion and Future Work
RAEplan demonstrates that even limited look-ahead significantly improves performance over purely reactive systems. By unifying the representation for acting and planning, the system avoids the consistency headaches of traditional architectures.
Future work will explore explicit robustness criteria, as the current model focuses primarily on efficiency. Additionally, the authors plan to incorporate temporal reasoning for complex time constraints and move from simulation to physical robot platforms.
Summary updated