Abstract
Gaze control requires the coordination of movements of both eyes and head to fixate on a target. Using our biologically constrained architecture for gaze control we show how the relationships between the coupled sensorimotor systems can be learnt autonomously from scratch, allowing for adaptation as the system grows or changes. Infant studies suggest developmental learning strategies, which can be applied to sensorimotor learning in humanoid robots. We examine environmental constraints for the learning of eye and head coupled mappings, and give results from implementations on an iCub robot. The results show the impact of these constraints and how they can be overcome to benefit the development of fast, cumulative, on-line learning of coupled sensorimotor systems.
Introduction
In the field of developmental robotics, the goal is to create systems that can adapt to changing bodies and environments, much like human infants. A fundamental challenge in this domain is gaze control: the ability to coordinate eye and head movements to fixate on a visual target. For a robot starting with no prior knowledge, the search space for these coordinated movements is immense. This paper investigates how internal biological constraints and external environmental factors shape the autonomous learning of these sensorimotor mappings on an iCub humanoid robot.
Methods: A Biologically Constrained Architecture
The researchers developed an architecture that mimics the neural structures found in biological systems. Central to this approach are 2D topographic neural maps that represent various sensorimotor spaces, including retinal positions and motor commands for the eyes and neck.
The learning process is driven by "motor babbling," where the robot performs random movements. When a movement successfully centers a target on the retina, a link is established between the sensory input and the motor command. These links form the basis of the topographic maps used for future saccades.
Developmental Constraints
The study examined two types of constraints:
- Internal (Type A): Simulating infant physiology, such as initially locking the robot's head to mimic weak neck muscles.
- Environmental (Type B): Varying the visual complexity of the surroundings, specifically comparing environments with a single target versus three targets.
Results and Discussion
The experiments revealed that constraints often act as a feature rather than a limitation. By initially restricting head movement, the robot is forced to master eye-only saccades, reducing the initial complexity of the learning task.
Overcoming Environmental Limitations
A key finding was the interaction between internal and environmental constraints. In a "boring" environment with only one target, eye-map learning was limited, reaching less than 50% coverage. However, once the head constraint was released, the robot's ability to move its head allowed it to "discover" new parts of its environment. This led to a dramatic surge in map coverage, eventually surpassing the coverage achieved in target-rich environments where the head remained locked.
Performance Metrics
In a three-target environment, the robot achieved nearly 80% map coverage. Testing showed that maps learned in these stimulating environments were significantly more accurate, achieving successful single-step fixation in up to 76% of trials, compared to a range of 34-64% in single-target scenarios.
Conclusion
This research demonstrates that "staged" learning—where physical and environmental restrictions are gradually lifted—enables humanoid robots to autonomously acquire complex coordination skills. The results suggest that internal biological constraints can compensate for poor learning environments by facilitating broader exploration. Future work will explore autonomous mechanisms for constraint release, allowing the robot to transition between developmental stages based on its own learning progress.
Summary updated