Unsupervised learning and temporal context to recall complex robot trajectories

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Robot trajectory learning is complex, especially with repeated or shared states.
• Existing models struggle with ambiguities during trajectory recall.
• Encoding spatial and temporal features is crucial for accurate trajectory learning.

Purpose of the Study:

• To propose an unsupervised neural network model for learning and recalling complex robot trajectories.
• To address ambiguities arising from repeated or shared states in robot arm configurations.
• To evaluate the model's learning, recall, and robustness across various trajectory datasets.

Main Methods:

• Utilized an unsupervised neural network with two synaptic weight groups trained via competitive and Hebbian learning.
• Implemented mechanisms including local/global context units, disabled neurons, and redundancy to handle state ambiguities.
• Simulated the model on diverse robot trajectory datasets to assess performance.

Main Results:

• The network successfully learned and recalled complex robot trajectories, including those with repeated or shared states.
• The proposed mechanisms effectively resolved ambiguities encountered during trajectory recall.
• The model demonstrated robustness and accurate reproduction of current and next states.

Conclusions:

• The developed unsupervised neural network offers a robust solution for learning and recalling complex robot trajectories.
• The model's ability to handle state ambiguities enhances its applicability in real-world robotics.
• Further simulations confirmed the model's effectiveness in learning, recall, and robustness evaluations.