Environment Predictive Coding for Visual Navigation
Authors: Santhosh Kumar Ramakrishnan, Tushar Nagarajan, Ziad Al-Halah, Kristen Grauman
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on the real-world scanned 3D environments of Gibson and Matterport3D show that our method obtains 2 6 higher sample-efficiency and up to 57% higher performance over standard imagerepresentation learning. |
| Researcher Affiliation | Collaboration | Santhosh K. Ramakrishnan1,2 Tushar Nagarajan1,2 Ziad Al-Halah1 Kristen Grauman1,2 1The University of Texas at Austin 2Facebook AI Research |
| Pseudocode | No | The paper describes the method using prose and diagrams but does not include formal pseudocode blocks or algorithms. |
| Open Source Code | Yes | Code and pre-trained models are publicly available: https://vision.cs.utexas.edu/projects/epc/ |
| Open Datasets | Yes | We perform experiments on the Habitat simulator (Savva et al., 2019b) with Matterport3D (Chang et al., 2017) and Gibson (Xia et al., 2018), two challenging and photorealistic 3D datasets |
| Dataset Splits | Yes | This results in 5,047 videos per agent, which we divide into an 80-20 train/val split (i.e., 2M training frames). We perform interactive RL training on 61 MP3D train scenes. We evaluate the learned policies on 11 val and 18 test scenes in MP3D, and 14 val scenes in Gibson. |
| Hardware Specification | No | The paper does not specify the exact hardware (e.g., GPU models, CPU types) used for running the experiments. It only mentions 'All models are trained in Py Torch (Paszke et al., 2019) with DD-PPO (Wijmans et al., 2020) for 13M-15M frames with 60 parallel processes.' |
| Software Dependencies | No | The paper mentions 'Py Torch (Paszke et al., 2019)' and 'DD-PPO (Wijmans et al., 2020)' but does not provide specific version numbers for these software components (e.g., PyTorch 1.9, Python 3.8). |
| Experiment Setup | Yes | We detail the list of hyperparameter choices for different tasks and models in Tab. A1. For ANS, we use 4 PPO mini-batches, 4 PPO epochs (not 2), and entropy coefficient of 0.003. |