Improving Embodied Instruction Following with Deterministic Methods

We propose a modified module structure of FILM, which increases its performance on ALFRED benchmark. Despite methods like imitation learning or reinforcement learning can be used, we will show it is possible to improve the performance by simply modifying the deterministic policy module. our approach aims to use the semantic segmented visual observation when navigating to the target location. Additionally, we propose a method of adaptive decision making for active search. Through active search, the agent can search for the target object by opening closed receptacles. Also, we provide some minor adjustments which helps the agent overcome some semantic segmentation errors. We evaluate the modified FILM with the ALFRED benchmark and achieved the performance of 25.37% with valid seen split and 20.46% with valid unseen split, thus succeeding in improving the performance of the original architecture with a small margin. Additionally, the modified FILM showed a significant improvement in the result of task type "Cool & Place" with a large margin (8.82%), which shows that active search is effective.