Stubborn: A Strong Baseline for Indoor Object Navigation

Grounding Implicit Goal Description for Robot Indoor Navigation via Recursive Belief Update

“…Different exploration strategies have been proposed based on these maps such as frontier-based search [17], maximizing exploration area [38] and object location prediction [6], [18]. Yet, recent work [39] observes that these maps are still insufficient for efficient object-oriented exploration. The other line of work implicitly learns representations of the environment in an end-to-end manner.…”

“…Once the target object is detected, the agent navigates to the target using the map or a PointNav policy. FBE [17] is a classical frontierbased exploration approach which builds an occupancy map and navigates to the nearest map frontiers; ANS [38] is a neural SLAM based exploration policy trained by RL to maximize area coverage; SemExp [2] builds a semantic map to predict the next goal trained by RL; PONI [6] is similar to SemExp [2] but uses behavior cloning for training; and STUBBORN [39] improves SemExp [2] by accumulating semantic segmentation results across steps for object recognition and utilizing a multi-scale occupancy map for path planning. Different from explicit mapping approaches which require hand-crafted map design and engineering, the implicit representation based methods [3], [4], [9], [5] train end-to-end models to predict actions, in particular the recurrent neural network.…”