Knowledge Guided Neural Network Task Planning for Generalization to Unseen Robotic Environments

This paper proposes the use of Long Short-Term Memory (LSTM) neural network architecture for High-level task action and location planning in robotic manipulation and navigation environments. We evaluated flat feedforward networks, Gated Recurrent Units (GRUs), and Long Short-Term Memory (LSTM) networks for predicting both actions and target locations and compared their performance to symbolic planners based on the Planning Domain Definition Language (PDDL). Experimental results demonstrate that LSTMs achieved a 95% success rate in manipulation tasks, significantly outperforming GRUs$(76 \%)$, and demonstrated strong generalization with$91\%$accuracy in unseen environments. Furthermore, LSTMs reduced planning latency by approximately$3.5 \times(0.031 \mathrm{s}$vs 0.109 s) compared to the PDDL baseline, validating its suitability for real-time robotic operation.