Local Manifold Approximation and Projection for Manifold-Aware Diffusion Planning
Authors: Kyowoon Lee, Jaesik Choi
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we present experimental results showing that augmenting prior diffusion planners with Lo MAP improves planning performance across a variety of offline control tasks. Specifically, we demonstrate (1) that Lo MAP effectively mitigates manifold deviation and filters out artifact trajectories, (2) that it further enhances planning performance when integrated into diffusion planner, and (3) that Lo MAP, as a plug-and-play module, can be seamlessly incorporated into hierarchical diffusion planners, enabling successful planning in the challenging Ant Maze domain. Additional details regarding our experimental setup and implementation are provided in Appendix E. |
| Researcher Affiliation | Academia | 1Korea Advanced Institute of Science and Technology (KAIST) 2INEEJI. Correspondence to: Jaesik Choi <EMAIL>. |
| Pseudocode | Yes | Pseudocode for the manifold-aware planning method is provided in Algorithm 1. Algorithm 1 Manifold-Aware Diffusion Planning... Algorithm 2 Lo MAP(τ i 1, {τ 0 n}, k) |
| Open Source Code | Yes | Codes are available at github.com/leekwoon/lomap. |
| Open Datasets | Yes | We validate our approach on standard offline reinforcement learning benchmarks that involve challenging long-horizon planning... We evaluate Lo MAP-incorporated planners on Mu Jo Co locomotion tasks (Fu et al., 2020), a standard benchmarks for assessing performance on heterogeneous, varying-quality datasets. The Ant Maze tasks (Fu et al., 2020) pose a substantial challenge due to high-dimensional state and action spaces, long-horizon navigation objectives, and sparse rewards. |
| Dataset Splits | No | For the high-level and low-level planners, we follow Chen et al. (2024) and train each component separately using trajectory segments randomly sampled from the D4RL offline dataset. This statement indicates the source and sampling method but does not specify explicit training, validation, or test split percentages or detailed methodology for reproducibility. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU specifications, or memory amounts used for running the experiments. It only discusses the experimental setup and implementation details for the software and models. |
| Software Dependencies | No | We build on the Diffuser framework (Janner et al., 2022)... we employ an inverted file (IVF) index from the Faiss library (Douze et al., 2024)... While it mentions the Diffuser framework and the Faiss library, it does not specify any version numbers for these or other software components, which is necessary for reproducible software dependencies. |
| Experiment Setup | Yes | Planning horizons. For Maze2D and Multi2D tasks, the planning horizon is 128 in U-Maze, 256 in Medium, and 256 in Large. For Mu Jo Co locomotion tasks, the horizon is 32, and for Ant Maze it is 64. Diffusion steps. We use 256 steps for the diffusion process in Maze2D//Multi2D Large and Medium, 128 in Maze2D//Multi2D U-Maze, and 20 in other environments. Guidance scales. For Ant Maze tasks, we select the guidance scale ω from the set {5.0, 3.0, 1.0, 0.1, 0.01, 0.001}. In Mu Jo Co locomotion tasks, we select ω from {0.3, 0.2, 0.1, 0.01, 0.001, 0.0001} during planning. Local manifold approximation. We tune the number of neighbors k {5, 10, 20} in our local manifold approximation procedure. Hierarchical Diffuser in Ant Maze. For the high-level and low-level planners, we follow Chen et al. (2024) and train each component separately using trajectory segments randomly sampled from the D4RL offline dataset. Specifically, the high-level planner generates state-space trajectories with a planning horizon of 226 and temporal jumps of 15. |