Hierarchical Abstraction for Combinatorial Generalization in Object Rearrangement

Authors: Michael Chang, Alyssa Li Dayan, Franziska Meier, Thomas L. Griffiths, Sergey Levine, Amy Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We compare NCS to several offline RL baselines and ablations on two rearrangement environments and find a significant gap in performance between our method and the next best method. Table 1: This table compares NCS with various baselines in the complete and partial evaluation settings of block-rearrange and robogym-rearrange.
Researcher Affiliation Collaboration Michael Chang , Alyssa L. Dayan, Franziska Meier, Thomas L. Griffiths, Sergey Levine, Amy Zhang work done as an intern at Meta AI. Correspondence to: mbchang@berkeley.edu and amyzhang@meta.com
Pseudocode Yes Algorithm 1 Building the Graph; Algorithm 2 Action Selection
Open Source Code No The paper states 'A step-by-step explanatory video of our method can be found in the supplementary material.' but does not explicitly provide or promise open-source code for the described methodology.
Open Datasets Yes The objects are uniformly sampled from a set of 94 meshes consisting of the YCB object set Calli et al. (2015) and a set of basic geometric shapes
Dataset Splits No The paper mentions 'The experiences buffer consists of 5000 trajectories showing 4 objects. We evaluate on 4-7 objects for 100 episodes across 10 seeds.' This describes the data used and evaluation setup, but not explicit train/validation/test splits with percentages or counts.
Hardware Specification No The paper mentions 'simulated rearrangement tasks' and references 'Mujoco: A physics engine for model-based control', but it does not specify any hardware used for running these simulations or training the models.
Software Dependencies No The paper cites 'Pybullet, a python module for physics simulation for games, robotics and machine learning. 2016.' and refers to 'JAXRL: Implementations of Reinforcement Learning algorithms in JAX, 10 2021.' but does not list specific version numbers for the key software components used in their own implementation.
Experiment Setup No The paper mentions 'The experiences buffer consists of 5000 trajectories showing 4 objects. We evaluate on 4-7 objects for 100 episodes across 10 seeds.' This provides some details about the evaluation setup but does not include specific hyperparameters (e.g., learning rate, batch size, number of epochs) or detailed system-level training settings.