General Evaluation for Instruction Conditioned Navigation using Dynamic Time Warping
This work addresses evaluation challenges in instruction-conditioned navigation, an incremental improvement for researchers and developers in robotics and AI.
The paper tackled the problem of evaluating instruction-conditioned navigation agents by addressing flaws in existing metrics, proposing the nDTW and SDTW metrics based on Dynamic Time Warping, which correlate better with human judgments and improve agent performance on R2R and R4R datasets, with SDTW showing superiority on R4R.
In instruction conditioned navigation, agents interpret natural language and their surroundings to navigate through an environment. Datasets for studying this task typically contain pairs of these instructions and reference trajectories. Yet, most evaluation metrics used thus far fail to properly account for the latter, relying instead on insufficient similarity comparisons. We address fundamental flaws in previously used metrics and show how Dynamic Time Warping (DTW), a long known method of measuring similarity between two time series, can be used for evaluation of navigation agents. For such, we define the normalized Dynamic Time Warping (nDTW) metric, that softly penalizes deviations from the reference path, is naturally sensitive to the order of the nodes composing each path, is suited for both continuous and graph-based evaluations, and can be efficiently calculated. Further, we define SDTW, which constrains nDTW to only successful paths. We collect human similarity judgments for simulated paths and find nDTW correlates better with human rankings than all other metrics. We also demonstrate that using nDTW as a reward signal for Reinforcement Learning navigation agents improves their performance on both the Room-to-Room (R2R) and Room-for-Room (R4R) datasets. The R4R results in particular highlight the superiority of SDTW over previous success-constrained metrics.