Distilling Morphology-Conditioned Hypernetworks for Efficient Universal Morphology Control

Authors: Zheng Xiong, Risto Vuorio, Jacob Beck, Matthieu Zimmer, Kun Shao, Shimon Whiteson

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments aim to answer the following questions: (1) Can Hyper Distill achieve good performance on both training and unseen test robots? How does it compare to other methods w.r.t. performance and efficiency at inference time? (Section 4.2) (2) How do different algorithmic and architecture choices in Hyper Distill influence its training and generalization performance? (Section 4.3)
Researcher Affiliation Collaboration 1Department of Computer Science, University of Oxford, Oxford, UK 2Huawei Noah s Ark Lab, London, UK.
Pseudocode No The paper does not contain any explicitly labeled pseudocode or algorithm blocks. It describes the methods in prose and with diagrams.
Open Source Code Yes The code is publicly available at https://github.com/Master Xiong/ Universal-Morphology-Control.
Open Datasets Yes We experiment on the UNIMAL benchmark (Gupta et al., 2021) built upon the Mujoco simulator (Todorov et al., 2012), which includes 100 training robots and 100 test robots with diverse morphologies
Dataset Splits No The paper mentions 100 training robots and 100 test robots, and an augmented set of 1000 PD robots for distillation. It does not explicitly define a separate validation split or subset for hyperparameter tuning or model selection in the main text.
Hardware Specification No The paper mentions that "The experiments were made possible by a generous equipment grant from NVIDIA." However, it does not provide specific details such as GPU models (e.g., A100, V100), CPU models, or memory specifications.
Software Dependencies No The paper mentions using the "Mujoco simulator" and "Adam" optimizer, but it does not specify version numbers for these or any other software components (e.g., Python, PyTorch, TensorFlow, etc.) used in the experiments.
Experiment Setup Yes The distillation process runs for 150 epochs, with a mini batch size of 5120. We use Adam with a learning rate of 0.0003, and clip the gradient norm to 0.5. For Hyper Distill, we apply dropout to context embedding with p = 0.1.