A Unified Cognitive Architecture for Physical Agents

Pat Langley, Dongkyu ChoiAAAI 2006

Introduction

In the pursuit of Artificial General Intelligence (AGI), the field of AI has frequently struggled with "fragmentation." Traditional approaches often isolate specific functions—such as computer vision, path planning, or reinforcement learning—into discrete modules that are difficult to integrate into a cohesive whole. This paper presents ICARUS, a cognitive architecture designed specifically for physical agents that operates in dynamic environments. Developed by Pat Langley and Dongkyu Choi, ICARUS seeks to unify perception, inference, execution, and learning within a single hierarchical framework, grounding high-level reasoning in raw sensory data.

The ICARUS Framework: Memory and Inference

ICARUS is built upon several interlocking memory systems that allow an agent to navigate the world and solve problems. At its core are two primary long-term memory structures:

  • Conceptual Memory: This stores the agent’s knowledge about the world using a formalism similar to PROLOG. On every cognitive cycle, ICARUS performs bottom-up inference, matching raw "percepts" from a perceptual buffer against these stored concepts to generate a "belief memory." This ensures that the agent's internal state is always grounded in its immediate physical surroundings.
  • Skill Memory: Skills are organized hierarchically, representing the agent's procedural knowledge. A top-down process evaluates the agent’s current goals and beliefs to select an appropriate "skill path." This allows for reactive, closed-loop control; because the agent re-evaluates its path every cycle, it can respond fluidly to changes in a dynamic environment.

Problem-Solving via Means-Ends Analysis

While ICARUS can execute routine skills reactively, it is also designed to handle novel situations where no pre-existing skill applies. When the agent encounters such an "impasse," it triggers a problem-solving module based on Means-Ends Analysis (MEA).

In this mode, the agent chains backward from its primary goal to identify subgoals that can be achieved using its current repertoire of skills. This deliberate reasoning process allows the agent to navigate complex scenarios that it has not been explicitly programmed to handle, effectively bridging the gap between reactive execution and high-level planning.

Structural Learning: From Deliberation to Routine

One of the most significant features of ICARUS is its approach to learning. Rather than treating learning as a separate training phase, ICARUS views it as a byproduct of successful problem-solving. When the MEA module successfully reaches a goal, the architecture performs "structural learning."

ICARUS "compiles" the sequence of subgoals and conditions used during the problem-solving episode into a new, permanent skill clause. This form of "learning by doing" transforms slow, deliberate reasoning into fast, routine execution. Over time, the agent becomes more efficient, as it no longer needs to invoke the computationally expensive problem-solver for tasks it has previously mastered.

Empirical Evaluation: The Driving Domain

The architecture was primarily evaluated in a simulated in-city driving environment featuring buildings, intersections, and other vehicles. The agent’s task was to navigate the vehicle, manage lane changes, and execute turns to reach a specific destination.

Initially, the agent possessed only 19 primitive skills. However, through active task performance and problem-solving, it successfully extracted five new high-level skill clauses. These newly acquired skills allowed the agent to perform complex maneuvers, such as lane changes, in future trials without needing to resort to backward chaining. Beyond driving, the authors note that ICARUS has demonstrated generality by successfully performing symbolic tasks like multi-column subtraction and the Tower of Hanoi.

Insights and Future Directions

ICARUS highlights several key principles for cognitive architecture:

  1. Grounding: All cognition must be rooted in perception and action.
  2. Skill-Concept Unity: Every skill is linked to a concept it aims to achieve, unifying "knowing that" with "knowing how."
  3. Hierarchy: Both beliefs and skills benefit from a hierarchical organization, allowing complex behaviors to be composed of simpler primitives.

Despite its successes, the architecture has limitations. Its current search mechanism is more systematic than human memory allows, and it lacks episodic memory for specific past events. Furthermore, the system currently only learns from successes, whereas human learning often stems from failure. Future work aims to address these gaps, including the development of social cognition mechanisms to better predict the intentions of other agents in shared environments.

Summary updated

BibTeX
@inproceedings{langley2006unified,
  author = {Pat Langley and Dongkyu Choi},
  title = {{A Unified Cognitive Architecture for Physical Agents}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure