Learning Latent Dynamic Robust Representations for World Models

Authors: Ruixiang Sun, Hongyu Zang, Xin Li, Riashat Islam

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical evaluation demonstrates significant performance improvements over existing methods in a range of visually complex control tasks such as Maniskill (Gu et al., 2023) with exogenous distractors from the Matterport environment.
Researcher Affiliation Collaboration 1Beijing Institute of Technology, China 2Dream Fold AI, Canada.
Pseudocode Yes Algorithm 1 HRSSM
Open Source Code Yes Our code is avaliable at https://github.com/ bit1029public/HRSSM.
Open Datasets Yes We perform our experiments in three distinct settings: i) a set of Mu Jo Co tasks (Todorov et al., 2012) provided by Deepmind Control(DMC) suite (Tassa et al., 2018), ii) a variant of Deep Mind Control Suite where the background is replaced with grayscale natural videos from Kinetics dataset (Kay et al., 2017), termed as Distracted Deep Mind Control Suite (Zhang et al., 2018), and iii) a benchmark based on the Maniskill2 (Gu et al., 2023), enhanced with realistic images of human homes (Chang et al., 2017) as backgrounds and was introduced in (Zhu et al., 2023).
Dataset Splits No The paper describes experiments conducted in various environments but does not provide explicit training, validation, or test dataset splits in terms of percentages or sample counts. Data is typically generated through interaction with the environment in RL.
Hardware Specification Yes We compare the wall-clock traning time of our method and Dreamer V3 in the Realistic Maniskill environment, with the use of a sever with NVidia A100SXM4 (40 GB memory) GPU.
Software Dependencies No The paper mentions using an 'unofficial open-sourced pytorch version of Dreamer V3(NM512, 2023)', but it does not specify the version numbers for PyTorch or other key software components used for reproducibility.
Experiment Setup Yes Table 3. Our model’s hyperparameters, which are the same across all tasks in DMControl and Realistic Maniskill. This table lists various hyperparameters such as Replay capacity (FIFO) 10^6, Batch size B 16, Batch length T 64, Learning rate 10^-4, Mask ratio 50%, Cube spatial size h w 10 10, etc.