MoVie: Visual Model-Based Policy Adaptation for View Generalization

Authors: Sizhe Yang, Yanjie Ze, Huazhe Xu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform extensive experiments on 7 robotic manipulation tasks (Adroit hand [20] and x Arm [7]) and 11 locomotion tasks (DMControl suite [30])), across the proposed 4 view generalization settings, totaling 18 4 configurations. In this section, we investigate how well an agent trained on a single fixed view generalizes to unseen views during test time.
Researcher Affiliation Collaboration Sizhe Yang12 , Yanjie Ze13 , Huazhe Xu415 1Shanghai Qi Zhi Institute, 2University of Electronic Science and Technology of China, 3Shanghai Jiao Tong University, 4Tsinghua University, 5Shanghai AI Lab Equal contribution
Pseudocode No No dedicated 'Pseudocode' or 'Algorithm' block was found. The paper provides network architecture descriptions in code-like format in Appendix A, but not a full algorithm for the overall method.
Open Source Code Yes Code and videos are available at yangsizhe.github.io/Mo Vie.
Open Datasets Yes Our test platform consists of 18 tasks from 3 domains: 11 tasks from DMControl [30], 3 tasks from Adroit [20], and 4 tasks from x Arm [7].
Dataset Splits No The paper describes training and testing phases and data collection during evaluation, but does not specify explicit training/validation/test dataset splits with percentages or counts, which is common in supervised learning.
Hardware Specification Yes One seed of our experiments could be run on a single 3090 GPU with fewer than 2GB and it takes 1 hours for test-time training.
Software Dependencies No The paper mentions using 'official implementation of TD-MPC [11] and Mo Dem [8]' but does not specify version numbers for general software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes We present relevant hyperparameters during both training and test time in Table 10 and Table 11.