Geometry-aware RL for Manipulation of Varying Shapes and Deformable Objects

Authors: Tai Hoang, Huy Le, Philipp Becker, Vien A Ngo, Gerhard Neumann

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results demonstrate that the proposed Heterogeneous Equivariant Policy (HEPi) outperforms both Transformer-based and pure EMPN baselines, particularly in complex 3D manipulation tasks. HEPi s integration of equivariance and explicit heterogeneity modelling improves performance in terms of average returns, sample efficiency, and generalization to unseen objects.
Researcher Affiliation Collaboration Tai Hoang1 , Huy Le1,2, Philipp Becker1, Ngo Anh Vien2, Gerhard Neumann1 1Autonomous Learning Robots, Karlsruhe Institute of Technology 2Bosch Center for Artificial Intelligence
Pseudocode No The paper describes the methodology using mathematical equations and prose. It does not include a distinct section labeled "Pseudocode" or "Algorithm", nor does it present structured code-like blocks.
Open Source Code No Our project page is available here.
Open Datasets Yes We then introduce a novel task, Rope-Shaping, which increases complexity by requiring the rope to form a specific shape (a W from the LASA dataset (Khansari-Zadeh & Billard, 2011)) to a desired orientation.
Dataset Splits Yes Finally, we evaluate the generalization of these models to unseen objects on two rigid tasks: rigid-sliding and rigid-insertion. Both tasks are trained on subsets of objects one (plus), two (plus, star), and three (plus, star, pentagon) and tested on the remaining objects.
Hardware Specification Yes All experiments were conducted on a machine equipped with an NVIDIA A100 or an NVIDIA H100 GPU.
Software Dependencies No We utilized the Torch RL framework (Bou et al., 2023) for the implementation of PPO and TRPL algorithms, and Py G (Py Torch Geometric) (Fey & Lenssen, 2019) for handling the graph-based structure. The Transformer architecture was implemented using the torch.nn.Transformer Encoder and torch.nn.Transformer Encoder Layer packages from Py Torch (Paszke et al., 2017).
Experiment Setup Yes We presents the hyperparameters used across all policy models (HEPi, EMPN, and Transformer) for all the tasks in Table 4. Table 5: Hyperparameters for Rigid Environments Table 6: Hyperparameters for Deformable Environments