IPAB Workshop-13/08/2020

Abstract: We propose an efficient learning-based policy for controlling an intelligent vehicle to negotiate obstacles, climb stairs and handle slippage, while most self-driving research focus on navigation and path planning. Due to uncertainties in real-world scenarios, eg obstacle and slippage, many traffic accidents happen which are challenging for self-driving vehicles to resolve. Therefore, closed-loop feedback control plays an important role in improving the robustness and resilience. We proposed the use of a long-short-term-memory (LSTM) neural network to learn reactive responsive control policies from human demonstrations. Our algorithm can directly learn robust control policies from datasets collected from a few real demonstrations. From the demo, the proposed controller learns the necessary skills for obstacle-negotiation and stair-climbing, including failure-recovery and some necessary corrective actions to handle the slippage. Furthermore, our proposed method can deal with non-optimal demos, since collecting enough amounts of suitable demos would be time-consuming and sometimes very difficult on real robotic systems.