ScottyActivity: Mixed Discrete-Continuous Planning with Convex Optimization
Authors: Enrique Fernandez-Gonzalez, Brian Williams, Erez Karpas
JAIR 2018 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 8. Experimental Results: In this section we evaluate the scalability of our planner in both synthetic domains and real expressive robotic scenarios. First, we use synthetic domains to illustrate why maintaining both continuous time and continuous control variables, as Scotty Activity does, is essential to plan efficiently over long horizons. We then present in Section 8.2 three robotic domains that Scotty Activity is designed to solve efficiently and show the scalability of our approach in these problems. Finally, we present in Section 8.3 a comparison of Scotty Activity against a mixed-integer solution (MIP). |
| Researcher Affiliation | Academia | Enrique Fern andez-Gonz alez EMAIL MIT Computer Science and Artificial Intelligence Laboratory ... Brian Williams EMAIL MIT Computer Science and Artificial Intelligence Laboratory ... Erez Karpas EMAIL Technion Israel Institute of Technology |
| Pseudocode | Yes | Algorithm 1: Get-Activities ... Algorithm 2: Test-Consistency ... Algorithm 3: Scotty-Plan-Ehc ... Algorithm 4: Scotty-Plan-Obj-Ehc |
| Open Source Code | No | The paper does not provide an explicit statement about making the source code publicly available or a link to a code repository for the methodology described. It mentions that POPCORN's authors "could not share the planner binary with us", implying limited access to external tools, but does not state the availability of their own planner's code. |
| Open Datasets | Yes | Appendix D. Benchmark Domains: In this appendix we provide the PDDL sources of some of the instances of the benchmark domains described in Section 8.2. |
| Dataset Splits | No | The paper evaluates the planner on specific problem instances within different domains (AUV, ROV, Air Refueling) provided as PDDL sources in Appendix D. The concept of training/test/validation dataset splits, as typically applied in machine learning for model training and evaluation, does not apply to these planning problems. Instead, the paper uses various problem instances (e.g., AUV instance 3, ROV instance 6) as individual test cases. |
| Hardware Specification | Yes | In our tests, we used an Intel Core i7-3770 3.40 GHz processor, and Gurobi 7 as Scotty Activity s internal convex optimization solver. |
| Software Dependencies | Yes | In our tests, we used an Intel Core i7-3770 3.40 GHz processor, and Gurobi 7 as Scotty Activity s internal convex optimization solver. ... We hypothesize that this is due to the superior performance of the Gurobi solver compared to the solvers used by Scotty1 (CPLEX 12.4) and POPCORN (lpsolve 5.5). |
| Experiment Setup | Yes | Appendix C. Example Scenario in PDDL-S Syntax and Appendix D. Benchmark Domains provide detailed PDDL definitions for problem instances, including specific durations (e.g., ":duration ( and (>= ? duration 0 . 1 ) (<= ? duration 200) )" for 'navigate ROV'), control variable bounds (e.g., ":control variable vx s :bounds ( and (>= ? value 2.0) (<= ? value 2 . 0 ) )"), region definitions (e.g., ":region rov range :parameters (? x1 ?y1 ?x2 ?y2 ) :condition ( and (max distance ((? x1 ?y1 ) (? x2 ?y2 ) ) :d 10) ) )"), and objective function weightings (e.g., ":metric minimize (+ ( 0.1 ( total time ) ) ( 2.5 (norm sq ( vel ship ) ) ) ) )" for the ROV domain). These extensively define the specific parameters and constraints of the planning problems used in the experiments. |