This research provides a high-level overview of a multi-year effort to develop "cognitively adequate" robot architectures. Presented at the IJCAI 2022 workshop on Cognitive Aspects of Knowledge Representation, the work introduces a hybrid framework that bridges the gap between high-level symbolic reasoning and low-level sensorimotor execution.
The Representation Gap in Robotics
Modern robotics faces a fundamental "representation gap." Robots require commonsense reasoning to understand general rules and domain knowledge (e.g., "textbooks are usually found in libraries"). Simultaneously, they must navigate noisy sensor data and quantitative uncertainty (e.g., "there is a 90% probability this object is a book").
Most existing systems prioritize one over the other: purely symbolic systems are often too fragile for the real world, while purely data-driven models act as "black boxes" that lack transparency and require massive datasets. This paper proposes an architecture designed for Cognitive Adequacy, allowing a robot to reason, learn, and justify its actions to human collaborators in plain language.
The REBA Framework: Stepwise Iterative Refinement
The core of the proposed architecture is the principle of Stepwise Iterative Refinement. The system, often referred to as the Refinement-Based Architecture (REBA), views the robot’s world through three distinct roles:
- The Logician: Uses Answer Set Prolog (ASP) to perform non-monotonic reasoning. This handles high-level planning and default assumptions.
- The Statistician: Manages fine-resolution transition diagrams, handling probabilistic models of uncertainty in perception and actuation as the robot moves from planning to concrete action.
- The Explorer: Employs Interactive Learning to update the robot's knowledge base. When the robot encounters unexpected outcomes, it uses inductive or reinforcement learning to refine its internal axioms.
Experimental Validation
The architecture has been validated through both simulated and physical environments. Rather than relying on massive generic datasets, the system identifies "regions of interest" where symbolic reasoning fails, using those specific instances to train deep networks. This makes the learning process highly data-efficient.
- Physical Implementation: The Baxter robot was used for real-world validation of object manipulation and stability estimation.
- Simulated Environments: Used to test complex planning and the robot's ability to disambiguate human commands in cluttered scenes.
Key Insights and Explainable Agency
The "Zooming" Mechanism
The robot maintains a coarse-grained view for high-level planning and only "zooms" into fine-grained details when necessary for execution. This prevents the system from being overwhelmed by computational complexity in large domains.
Reasoning-Guided Learning
Symbolic reasoning acts as a filter for machine learning. If the robot cannot explain a scene using its current logic, it identifies that specific data point as a priority for training its neural networks.
Explainable Agency
Because the robot’s internal state is grounded in relational logic, it can answer "Why?" or "Why not?" questions. For example, it can explain that it did not pick up a specific object because it was physically blocked by another item, providing a clear trace of its decision-making process.
Results and Future Directions
The hybrid approach significantly improves task reliability—such as estimating object stability—compared to using deep networks in isolation. By constraining the search space with symbolic knowledge, the robot requires fewer labeled examples to learn new tasks.
While successful, challenges remain in reducing the manual engineering required to design the initial "sorted signature" (the list of objects and actions). Future work aims to develop "adaptive theories of intentions" to help robots better manage complex, multi-stage goals and persistence in dynamic environments.
Summary updated