Abhijeet Kishore Nayak

Modern vision-based robotic navigation systems often struggle with a lack of robustness, especially when operating in complex, unstructured, and dynamic human environments. This project aims to overcome that limitation by developing new frameworks for robust visual navigation driven by generative action policies. A key contribution is the architectural advancement of these policies to provide inherent robustness to unexpected obstacles while ensuring strict collision avoidance. Another central element of the project is language grounding, enabling the robot to interpret human instructions and adapt its action policies based on a deeper semantic understanding of its surroundings. This conditioning only enhances the robot's comprehension of the environment but also supports long-horizon planning. In addition, incorporating episodic memory modules is essential for generating efficient and context-aware actions. These memory components allow the robot to maintain a coherent internal representation of its environment and support safe, consistent operation. Together, these innovations lead to navigation policies that are safer, more accurate, and more interpretable. Ultimately, the goal is to build a deployable robotic system that navigates safely and adapts its behavior through a unified understanding of visual input, instruction cues, and persistent environmental memory.