Multi-modal Predicate Identification using Dynamically Learned Robot Controllers
Authors: Saeid Amiri, Suhua Wei, Shiqi Zhang, Jivko Sinapov, Jesse Thomason, Peter Stone
IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed method using two datasets in which a robot explored a set of objects using a variety of exploratory behaviors and sensory modalities, and show that, for both, our proposed MOMDP model outperforms baseline models in exploration accuracy and overall exploration cost. |
| Researcher Affiliation | Academia | Cleveland State University, Cleveland, OH 44115 USA Tufts University, Medford, MA 02155 USA The University of Texas at Austin, Austin, TX 78712 USA |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described. |
| Open Datasets | Yes | Two datasets of Sinapov14 and Thomason16 are used in the experiments, where Thomason16 has a much more diverse set of household objects and a larger number of predicates that arose naturally during human-robot interaction gameplay. [...] Sinapov14 Dataset: In this dataset, the robot explored 36 different objects using 11 prototypical exploratory behaviors: look, grasp, lift, shake, shake-fast, lower, drop, push, poke, tap, and press 10 different times per object [Sinapov et al., 2014b]. [...] Thomason16 Dataset: In this dataset, the robot, part of the Building-Wide Intelligence project [Khandelwal et al., 2017], explored 32 common household objects using 8 exploratory actions: look, grasp, lift, hold, lower, drop, push, and press. Each behavior was performed 5 times on each object. The dataset was originally produced for the task of learning how sets of objects can be ordered and is described in greater detail by [Sinapov et al., 2016]. |
| Dataset Splits | No | The paper mentions 'cross-validation on the training data' for predicate learning and 'cross-validation at the behavior level' for confusion matrices, but does not specify explicit validation dataset splits (e.g., percentages or sample counts) for the overall experimental setup. |
| Hardware Specification | No | The paper refers to robots used for data collection and interaction (e.g., 'a robot equipped with an arm', 'different robots' for datasets) but does not provide specific hardware details (like CPU/GPU models, memory, or cloud instances) used for running the computational experiments or training the models. |
| Software Dependencies | No | The paper references existing POMDP solvers and networks (e.g., VGG network) but does not specify version numbers for any software dependencies used in the experiments. |
| Experiment Setup | Yes | Currently, building the model takes almost no time, and we uniformly gave five seconds for policy generation using the model (same in all experiments). [...] Unless otherwise specified, r = 500 and r+ =500 in this paper. [...] Costs of other exploration actions come from the datasets used in this research, and are within the range of [0.5,22.0] (corresponding reward is negative), except that action ask has the cost of 100.0. γ is a discount factor, and γ = 0.99 in our case. |