Learning Active Camera for Multi-Object Navigation

Authors: Peihao Chen, Dongyu Ji, Kunyang Lin, Weiwen Hu, Wenbing Huang, Thomas Li, Mingkui Tan, Chuang Gan

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show our camera policy consistently improves the performance of multi-object navigation over four baselines on two datasets.
Researcher Affiliation Collaboration 1South China University of Technology, 2Pazhou Laboratory, 3MIT-IBM Watson AI Lab, 4UMass Amherst, 5Gaoling School of Artificial Intelligence, Renmin University of China, 6Information Technology R&D Innovation Center of Peking University, 7Key Laboratory of Big Data and Intelligent Robot, Ministry of Education
Pseudocode Yes Algorithm 1 Training method for active-camera agent.
Open Source Code No We will publish our code upon acceptance.
Open Datasets Yes We perform experiments on two photorealistic 3D indoor environments, i.e., Matterport3D [7] and Gibson [71].
Dataset Splits Yes We have included the training details in the supplemental materials.
Hardware Specification Yes We have included the details of the computational resources in the supplemental material.
Software Dependencies No The paper does not explicitly list specific software dependencies with version numbers.
Experiment Setup Yes For navigation and camera actions, a FORWARD action moves the agent forward by 0.25 meters and a TURN action turns by 30 . The maximum episode time step is 500... We set K 8, r 2.4, γ rˆ1.2 2.88 in heuristic module empirically... Reward scaling factors α and β are set to 10 and 1 respectively such that three reward terms are in the same order of magnitude at initialization.