Introduction
By the mid-1980s, a significant divergence had emerged between the fields of Artificial Intelligence (AI) and Robotics. AI researchers were primarily occupied with "abstracted worlds"—logical puzzles like chess or block-stacking—while robotics focused on the physical realities of kinematics, dynamics, and hardware control. This paper identifies a critical "missing link" in this evolution: system integration.
The author argues that AI programs often operated under the assumption of perfect execution, while robotic controllers lacked the semantic understanding necessary for intelligent behavior. To address this, the paper introduces COSMOS, an integrated environment designed to bridge the gap between high-level symbolic planning and low-level physical action.
The COSMOS System Architecture
The core of the proposed solution is a unified framework where perception, cognition, and action are tightly coupled. The system is built upon a Lisp environment (ETALisp), facilitating interactive programming and symbolic manipulation for complex task planning.
The architecture employs a hierarchical approach to manage complexity:
- Abstracted Manipulator System: To shield high-level AI from the complexities of joint angles and torques, the system defines seven "semantic" functions (e.g.,
put-arm-reference,get-wrist-force). This provides a clean API for hardware interaction. - Hierarchical Robot Languages: The system utilizes AL/L, a manipulation-oriented language, alongside higher-level task-oriented layers capable of planning sequences based on geometric constraints.
- Environment Model Management: A central requirement for intelligence is maintaining a "digital twin" of the environment. COSMOS ensures this model stays synchronized with reality; for instance, if a robot drops an object, the system detects the inconsistency via vision and triggers a re-planning phase.
Multi-Window Vision and Real-Time Interaction
A standout technical contribution of the COSMOS system is its novel hardware-based vision processor. Moving away from traditional full-frame processing, the author proposes a "multi-window" approach. This system utilizes 16 independent "windows" or "attentions" that can be controlled in parallel. Each window can be independently configured for location, shape, and resolution.
This "visual attention" mechanism allows the robot to monitor multiple key features simultaneously—such as the end of a rope and a target ring—in real-time. This capability is essential for handling non-rigid objects, which do not have fixed, well-defined shapes and require constant monitoring of preconditions and postconditions.
Experimental Results: The Rope-Tying Task
To demonstrate the efficacy of the integrated approach, the author chose the complex task of rope manipulation. Unlike rigid blocks, rope handling requires constant visual feedback and adaptive planning.
The COSMOS system successfully performed a "hand-eye" task involving finding a rope, threading it through a ring using visual feedback, and ultimately tying a knot. The multi-window vision system proved vital, allowing the robot to track the rope's position and the ring's location in parallel. This experiment confirmed that a 32-bit minicomputer (the Eclipse MV/4000) could effectively host a Lisp-based "brain" capable of controlling multiple robotic arms and sensors in a unified framework.
Limitations and Future Directions
Despite the success of COSMOS, the author highlights several areas for future research:
- Primitive Planners: Current automated problem solvers struggle with the messy constraints of real-world environments.
- Ad-Hoc Strategies: Many manipulation behaviors still rely on specific, non-generalized strategies. The author calls for a formal "Theory of Manipulation."
- Hardware Scaling: Future systems would benefit from custom LSI (Large Scale Integration) chips to make multi-attention vision more compact and powerful.
- Model Consistency: Maintaining a perfect internal world model remains a significant challenge, particularly in the presence of sensor noise or unexpected environmental changes.
By integrating symbolic reasoning with real-time perception and control, COSMOS represents a foundational step toward creating robots that are not just tools, but intelligent agents capable of interacting with a complex, physical world.
Summary updated