Learning to Move with Affordance Maps
Authors: William Qi, Ravi Teja Mullapudi, Saurabh Gupta, Deva Ramanan
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach in the Viz Doom simulator, using large-scale randomly-generated maps containing a variety of dynamic actors and hazards. We show that learned affordance maps can be used to augment traditional approaches for both exploration and navigation, providing significant improvements in performance. |
| Researcher Affiliation | Academia | 1 Carnegie Mellon University 2 UIUC |
| Pseudocode | No | The paper describes its methods and pipeline using natural language and diagrams (Figure 1, Figure 2) but does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | No | As the proposed system contains a large number of moving parts and tunable hyper-parameters, we plan to release modular open-source code for the described system at the following location: https://github.com/wqi/A2L. |
| Open Datasets | No | The paper states: "We conduct our experiments within procedurally-generated Viz Doom maps created by the Oblige (Apted, 2017) level generator" and "We collect approximately 100k total samples across the 60 training maps in a self-supervised manner". While Viz Doom and Oblige are publicly known, the specific generated maps or the collected 100k samples that form the training dataset are not explicitly made publicly available with a link or citation for access. |
| Dataset Splits | No | The paper mentions 60 training maps and 15 test maps, and states "train the navigability module for 50 epochs using the collected dataset". However, it does not specify a separate validation split (e.g., percentages or counts) for the collected dataset used during training. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used to run the experiments. It only mentions using the Viz Doom simulator. |
| Software Dependencies | No | The paper mentions using a "ResNet-18-based (He et al., 2016) UNet (Ronneberger et al., 2015) architecture" for segmentation, and PPO (Schulman et al., 2017) for RL, but does not provide specific version numbers for any software libraries, frameworks, or environments used (e.g., Python version, TensorFlow/PyTorch version, specific Viz Doom API version). |
| Experiment Setup | Yes | We evaluate this approach using hyper-parameters identical to those proposed by the original authors, with the only exception being the addition of a new penalty in the reward that is scaled by the amount of damage received at each time step. We collect approximately 100k total samples across the 60 training maps in a self-supervised manner and train the navigability module for 50 epochs using the collected dataset; a ResNet-18-based (He et al., 2016) UNet (Ronneberger et al., 2015) architecture is employed for segmentation. A masked loss function Lmask = K LBCE(ˆy, y) based on binary cross-entropy is employed. |