Abstract
Traditional AI planning systems often struggle in real-world environments because their internal models of physics and causality are necessarily incomplete. This paper introduces Permissive Planning, a machine learning approach that shifts the focus from refining the world model to adjusting the planner's decision-making bias. By learning to select "permissive" plans—those that remain successful despite model inaccuracies—the system achieves high reliability in robotic tasks without the computational burden of building a perfect domain theory.
Introduction: The Gap Between Model and Reality
In classical AI, a planner operates by "projecting" the effects of actions within an internal model. However, the real world is fraught with complexities like friction, sensor noise, and unforeseen obstacles. A robot might calculate that lifting its arm 1.4 units is sufficient to clear a block, only to collide with it due to slight mechanical deviations or model simplifications.
Historically, researchers have addressed this through two extremes: refining the model until it accounts for every variable (the "qualification problem"), or abandoning planning for purely reactive behaviors. DeJong and Bennett propose a third way. They observe that a planner often has a "competence set"—a range of plans that should work according to its model. The failure often lies not in the model itself, but in the planner's choice of a non-robust plan from that set.
The Permissive Planning Principle
The core of this approach is the Permissive Planning Principle: when a plan fails, the system blames the planner’s choice rather than the underlying domain theory. It then adjusts its internal bias to favor more robust alternatives.
ISolves vs. ESolves
The authors distinguish between two types of solutions:
- ISolves (Internal Solutions): Plans that the internal model predicts will succeed.
- ESolves (External Solutions): Plans that actually succeed when executed in the physical world.
An ideal planner is "adequate" if every ISolution it generates is also an ESolution. Permissive planning achieves this adequacy by searching through a "bias space" to find parameters that maximize the likelihood of external success.
Plan Computation and Bias Adjustment
When a failure occurs, the system uses qualitative domain knowledge to explain the error. For instance, if a collision happens, the system recognizes that "higher is safer." It then prunes the space of possible biases to exclude the failed parameter.
As shown in Figure 6, the planner might initially select a performance value of 1.4 for an "Up" operator. After a failure and subsequent learning, the "permissive adjustment" pushes the selection toward a safer extreme, such as the "ceiling." This specific commitment increases the plan's robustness against model inaccuracies.
Experimental Evaluation
The approach was validated on two robotic testbeds:
- Gantry Robot Grasping: A 2D arm tasked with picking up randomly placed puzzle pieces. Without learning, the robot succeeded in only 2 out of 12 attempts. With permissive planning, success rose to 10 out of 12.
- Tilted Tray Orientation: A benchmark task involving orienting a rectangular object by tilting a tray. The permissive planner improved its success rate from 40% to 80% after just 20 repetitions.
Notably, the permissive planner achieved results comparable to complex stochastic (probabilistic) planners but with significantly lower computational overhead and fewer training examples because it avoided calculating exact probabilities for every possible outcome.
Conclusion
Permissive planning represents a paradigm shift in robotic reasoning. By accepting that internal models will always be imperfect, AI systems can focus on finding "permissive" regions of the state space where those imperfections do not lead to failure. While challenges remain regarding the "utility problem"—the cost of searching an ever-growing library of learned biases—this approach provides a practical bridge between high-level reasoning and low-level execution.
Summary updated