MultiON: Benchmarking Semantic Map Memory using Multi-Object Navigation

Authors: Saim Wani, Shivansh Patel, Unnat Jain, Angel Chang, Manolis Savva

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform a set of multi ON experiments to examine how a variety of agent models perform across a spectrum of navigation task complexities. Our experiments show that:
Researcher Affiliation Academia 1IIT Kanpur 2UIUC 3Simon Fraser University
Pseudocode No The paper provides architectural diagrams (Figure 1 and Figure 2) and describes the agent models and training process in text, but it does not include any formally presented pseudocode or algorithm blocks.
Open Source Code Yes https://shivanshpatel35.github.io/multi-ON/ (This link leads to a project page which then provides a direct GitHub link to the code: https://github.com/shivanshpatel35/multi-ON)
Open Datasets Yes We generate datasets for the 1-ON, 2-ON, and 3-ON tasks on the Matterport3D [11] scenes using the standard train/val/test split.
Dataset Splits Yes For each m-ON dataset the train split consists of 50,000 episodes per scene, and the validation and test splits each contain 12,500 episodes per scene.
Hardware Specification No The paper mentions training with '16 parallel threads' and indicates support from 'West Grid' and 'Compute Canada', which are high-performance computing organizations. However, it does not provide specific hardware details such as CPU/GPU models, memory, or cloud instance types used for the experiments.
Software Dependencies No The paper mentions the use of 'proximal policy optimization (PPO)' as a training algorithm but does not list any specific software dependencies or libraries with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes The agent is trained using the reward structure defined above using proximal policy optimization (PPO) [40]. All the agent models are trained for 40 million frames using 16 parallel threads. We use 4 mini-batches and do 2 epochs in each PPO update. We use rgoal = 3.0 and α = 0.01 for all experiments. The maximum number of steps is 2,500.