Learning Goal-Oriented Hierarchical Tasks from Situated Interactive Instruction

Authors: Shiwali Mohan, John Laird

AAAI 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experiments were conducted on a simulated environment that is faithful to the continuous sensor data and control policies the robot works with. We used the simulator to generate the reported results because the evaluation space is large and generating results with the real robot is extremely time consuming. A representative sample of the scenarios described in the paper were successfully run on the real robotic arm, achieving results that are consistent with those reported in the paper.
Researcher Affiliation Academia Shiwali Mohan and John E. Laird Computer Science and Engineering Division, University of Michigan 2260 Hayward Street, Ann Arbor, Michigan 48109-2121 {shiwali,laird}@umich.edu
Pseudocode Yes Algorithm 1 Executing a task. 1: procedure execute(state s, task t)...; Algorithm 2 Exploring the action space 1: procedure search(state s, desired d, actions C, policy π)...; Algorithm 3 Explaining instructions retrospectively 1: procedure explain-instructions(time n, task t)...
Open Source Code No The paper does not provide any statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets No The experiments were conducted on a simulated environment that is faithful to the continuous sensor data and control policies the robot works with. We used the simulator to generate the reported results because the evaluation space is large and generating results with the real robot is extremely time consuming.
Dataset Splits No A task trial consists of a series of episodes in each of which the agent is asked to execute the task with randomly generated parameters. Each episode begins in an initial state obtained by assigning random states to locations open/close(pantry), the arm (hold/ holds(o)), and arbitrarily placed objects on the workspace.
Hardware Specification No Rosie is embodied as a robotic arm (in Figure 1, left) that manipulates small foam blocks on a table-top workspace that simulates a kitchen. The experiments were conducted on a simulated environment that is faithful to the continuous sensor data and control policies the robot works with.
Software Dependencies No Our approach is implemented in Rosie (Mohan et al. 2012) a generally instructable agent developed using the Soar cognitive architecture (Laird 2012).
Experiment Setup Yes The exploration depth was set to 0 of this experiment. A sample of the results generated from the experiment is shown in Figure 2. ... At depth 0 the agent does not perform any exploration and relies completely on instructions. ... At depth 4, it discovers the solution and does not ask any child-queries.