Potential Based Diffusion Motion Planning

Authors: Yunhao Luo, Chen Sun, Joshua B. Tenenbaum, Yilun Du

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we first describe our environments and baselines in Section 4.1. Next, in Section 4.2, we discuss our experiments on the base environments and the motion refining algorithm. Following, in Section 4.3, we present the compositionality results by evaluating our motion planner on composite environments. Then, we describe the real world motion planning performance in Section 4.4.
Researcher Affiliation Academia 1Brown University 2MIT. Correspondence to: Yunhao Luo <yluo73@cs.brown.edu>.
Pseudocode Yes Algorithm 1 Compositional Potential Based Planning
Open Source Code Yes Project website at https://energy-based-model.github.io/potentialmotion-plan.
Open Datasets Yes Finally, we evaluate the effectiveness of our method on the real world ETH/UCY (Pellegrini et al., 2010; Lerner et al., 2007) dataset.
Dataset Splits No The paper describes training data and evaluation data (unseen environments) but does not explicitly provide details for a separate validation split, nor exact percentages or absolute sample counts for such a split.
Hardware Specification Yes Hardware:. For each of our experiments, we used 1 RTX 3090 GPU.
Software Dependencies Yes Software:. The computation platform is installed with Red Hat 7.9, Python 3.8, Py Torch 1.10.1, and Cuda 11.1
Experiment Setup Yes We provide detailed hyperparameters for training our model in Table 6. We do not apply any hyperparameter search nor learning rate scheduler. The training time of our model is approximately two days, but we observe that the performance is close to saturation within one day.