Curious Exploration via Structured World Models Yields Zero-Shot Object Manipulation

Authors: Cansu Sancaktar, Sebastian Blaes, Georg Martius

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our empirical evaluation, we analyze the performance of CEE-US on two different object manipulation environments to answer the questions: How much does the structural inductive bias introduced by GNNs help model learning and control? Does the free-play phase create rich interaction data that helps downstream task performance? Can we solve challenging manipulation tasks in a zero-shot manner?
Researcher Affiliation Academia Cansu Sancaktar Sebastian Blaes Georg Martius Max Planck Institute for Intelligent Systems Tübingen, Germany {cansu.sancaktar, sebastian.blaes, georg.martius}@tue.mpg.de
Pseudocode Yes Algorithm 1 CEE-US: Free Play in Intrinsic Phase, Algorithm 2 CEE-US: Zero-shot Generalization in Extrinsic Phase
Open Source Code Yes Code and videos are available at https://martius-lab.github.io/cee-us.
Open Datasets Yes Fetch Pick & Place Construction This is an extension of the Fetch Pick & Place environment [23] to more cubes [24] (Fig. 2).
Dataset Splits No The paper mentions "collected rollouts are added to the buffer" and "fixed dataset" in the context of training, and refers to "500k datapoints" and "600k datapoints" in Table 3 but does not specify explicit training/validation/test dataset splits with percentages or counts.
Hardware Specification No The paper states "Further details are in the Suppl. C" regarding compute, but the provided text does not include the content of Suppl. C, and no specific hardware models (GPU/CPU) are mentioned in the main body.
Software Dependencies No The paper mentions using "i CEM", "DDPG", "CQL", and the "d3rlpy library" but does not provide specific version numbers for any of these software dependencies.
Experiment Setup No The paper mentions the ensemble size (M=5) and the use of specific algorithms like i CEM, DDPG, and CQL, but refers to "Suppl. C" and "Suppl. C.5" for training details and hyperparameters, which are not provided in the main text.