Abstract
A fundamental challenge in autonomous robotics is the transformation of shifting sensory data into stable motor commands. This paper presents a hierarchical connectionist network that enables a robot to learn its own internal reference frames through experience. By utilizing intermediate invariant layers, the system effectively decouples target positions from the robot's own posture, leading to a significant improvement in reaching accuracy over flat architectures.
Introduction
Autonomous agents, whether biological or robotic, lack a fixed external reference frame. Sensory inputs, such as retinal images, change constantly as the agent moves its eyes or head. To interact accurately with the environment, the agent must translate these dynamic inputs into a stable internal representation. In 1988, Pabon and Gossard demonstrated that such a representation could be learned autonomously, rather than pre-programmed, by "closing the motor coordination loop" through self-supervised experience.
Methods
The study utilized a simulated robot with three hierarchical subsystems: a 2-degree-of-freedom (DOF) eye, a 2-DOF head, and a 2-DOF planar arm. The control mechanism was a hierarchical connectionist network trained via the delta rule to minimize motor signal prediction errors.
Key components of the architecture included:
- 2D Activation Maps: Gaussian distributions of activity in 2D arrays represented sensory and motor positions.
- Internal Invariant Layers: Two critical stages were introduced: the Target Position Map - Eye invariant (TPME) and the Target Position Map - Head invariant (TPMH). These layers maintain a constant target representation regardless of eye or head movement.
- Self-Supervised Learning: The robot learned by moving its arm randomly and using the resulting sensory feedback as input, with the original random motor signals serving as the reference target.
The system was trained over 10,000 iterations of self-generated movements to cover the workspace. Performance was evaluated based on the "global error," defined as the distance between the target and the arm's endpoint as a percentage of link length.
Key Results and Insights
The hierarchical approach proved essential for successful coordination. The system achieved a steady-state error of 8.4% of the arm link length. In contrast, removing the internal invariant layers caused the error to surge to approximately 50%, demonstrating a ten-fold improvement in accuracy attributable to the hierarchical structure.
The learning curves indicate asymptotic convergence, showing that the robot progressively masters its workspace. Furthermore, the authors highlight the biological relevance of this model, noting similarities between the TPME layer and the function of the posterior parietal cortex in primates. The distributed nature of the neural network also provides "graceful degradation," ensuring robustness against the loss of individual units.
Conclusion
This research underscores the necessity of hierarchical processing for autonomous spatial learning. By building internal invariant reference frames, robots can overcome the limitations of shifting sensory perspectives. While the model uses simplified 2D physics and specific activation distributions, it lays the groundwork for more complex 3D systems capable of adapting to physical changes or damage in real-time.
Summary updated