Autonomous Learning of Object-Centric Abstractions for High-Level Planning
Authors: Steven James, Benjamin Rosman, George Konidaris
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We first demonstrate our approach on a 2D crafting domain consisting of numerous objects where the agent learns a compact, lifted representation that generalises across objects. We then apply it to a series of Minecraft tasks to learn object-centric representations and object types directly from pixel data that can be leveraged to solve new tasks quickly. The resulting learned representations enable the use of a task-level planner, resulting in an agent capable of transferring learned representations to form complex, long-term plans. |
| Researcher Affiliation | Collaboration | Steven James & Benjamin Rosman University of the Witwatersrand Johannesburg, South Africa {steven.james,benjamin.rosman1}@wits.ac.za George Konidaris Brown University Providence RI 02912, USA gdk@cs.brown.edu |
| Pseudocode | Yes | We summarise our proposed approach in Figure 1 and provide a high-level description in the remainder of this section, but provide pseudocode and specific practical details in the appendix. ... Appendix I: PSEUDOCODE |
| Open Source Code | No | The paper provides a link to 'More results and videos' (https://sites.google.com/view/mine-pddl) but does not explicitly state that source code for the methodology is available there or elsewhere. |
| Open Datasets | No | The paper references environments like '2D crafting domain (Andreas et al., 2017)' and 'Minecraft task (Johnson et al., 2016)' and describes collecting its own data within these environments ('We collect data from a task by executing options uniformly at random'). It does not provide concrete access information (link, DOI, repository, or explicit statement of a pre-existing public dataset) for the specific datasets used in its experiments. |
| Dataset Splits | Yes | Then, for every other object in the environment, we compute the SVM’s accuracy when that object’s features are added to the SVM. Any object that increases the SVM accuracy is kept. Pseudocode for this procedure is outlined in Figure 9. ... We use a grid search with 3-fold cross validation to find the best bandwidth hyperparameter for each estimator. |
| Hardware Specification | No | The authors acknowledge the Centre for High Performance Computing (CHPC), South Africa, for providing computational resources to this research project. Computations were also performed using High Performance Computing Infrastructure provided by the Mathematical Sciences Support unit at the University of the Witwatersrand. This indicates the use of HPC resources but does not provide specific hardware details like GPU/CPU models or memory. |
| Software Dependencies | No | The paper mentions various algorithms and tools used, such as 'DBSCAN clustering algorithm (Ester et al., 1996)', 'SVM (Cortes & Vapnik, 1995) with Platt scaling (Platt, 1999)', 'kernel density estimator (Rosenblatt, 1956)', and the 'Malmo platform'. However, it does not specify version numbers for any of these software components or programming languages. |
| Experiment Setup | Yes | We downscale images to 160x120 and then convert the resulting images to greyscale. We apply principal component analysis (Pearson, 1901) to a batch of images collected from the different tasks and keep the top 40 principal components. ... Partitioning was achieved using DBSCAN with ϵ = 0.1. (ii) Preconditions were learned using an SVM with C = 4. (iii) The kernel density estimators used for the effects had bandwidth parameter 0.001. |