Mitigating Catastrophic Forgetting in Imitation Learning for Embodied AI using Progressive Neural Networks

Imitation learning is crucial for agent training to replicate expert behaviour, but it faces issues like catastrophic forgetting, where the model accommodates new tasks but forgets previously learned information.The research aims to develop a Progressive Neural Network and Elastic Weight Consolidation (PNN-EWC) approach that mitigates forgetting while enhancing task retention and adaptation.The PNN-EWC model is experimented with in the Webots simulation environment utilising the Franka Emika Panda robotic arm.This model integrates Progressive Neural Networks, which add new columns to the neural network for each task, and Elastic Weight Consolidation regularises updates to preserve learned tasks.The proposed PNN-EWC model achieved a 95% average success rate and a 6% forgetting rate in preventing catastrophic forgetting, highlighting significant improvements over traditional methods.This approach allows the Franka Emika Panda to continuously learn new tasks without losing previously acquired skills, making it effective for dynamic environments.