Semi-parametric topological memory for navigation
Authors: Nikolay Savinov, Alexey Dosovitskiy, Vladlen Koltun
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed system and relevant baselines on the task of goal-directed maze navigation in simulated three-dimensional environments. The proposed system outperforms baseline approaches by a large margin. Given 5 minutes of maze walkthrough footage, the system is able to build an internal representation of the environment and use it to confidently navigate to various goals within the maze. The average success rate of the SPTM agent in goal-directed navigation across test environments is higher than the best-performing baseline by a factor of three. Qualitative results and an implementation of the method are available at https://sites.google.com/view/SPTM. |
| Researcher Affiliation | Collaboration | Nikolay Savinov ETH Z urich Alexey Dosovitskiy Intel Labs Vladlen Koltun Intel Labs |
| Pseudocode | No | The paper describes the components and processes of SPTM verbally and with diagrams, but it does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Qualitative results and an implementation of the method are available at https://sites.google.com/view/SPTM. |
| Open Datasets | No | The paper describes generating training data by 'executing a random agent in the training environment' using a simulated Doom environment, but it does not provide concrete access (e.g., a link or citation) to the specific generated dataset used for training, nor does it claim the generated dataset is publicly available. |
| Dataset Splits | Yes | We used different mazes for training, validation, and testing. In addition, we created 3 mazes for validation and 7 mazes for testing. |
| Hardware Specification | No | The paper mentions the use of ResNet-18 as a base architecture and discusses the size of the networks, but it does not specify any particular hardware components like CPU or GPU models used for training or experimentation. |
| Software Dependencies | No | The paper mentions 'Keras (Chollet et al., 2015) and TensorFlow (Abadi et al., 2016)' as the frameworks used for implementation, but it does not provide specific version numbers for these software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | We train the networks R and L for a total of 1 and 2 million mini-batch iterations, respectively. We set this number to 2000 in what follows. When making visual shortcuts in the graph, we set the minimum shortcut distance Tℓ= 5 and the smoothing window size Tw = 10. The threshold values for waypoint selection are set to slocal = 0.7 and sreach = 0.95. The minimum and maximum waypoint distances are set to Hmin = 1 and Hmax = 7, respectively. |