ICAPS-24 System Demonstrations
All demo extended abstracts can be found here
The League of Robot Runners Competition: Goals, Designs, and Implementation
![](demo1.png)
Multi-Agent Path Finding (MAPF) is an important practical problem found in many application settings, from logistics and transportation to robotics and automation. In this paper, we introduce the goals, designs and implementations of the League of Robot Runners (LoRR), a competition to foster and advance this research area. LoRR aims to identify the core challenges for solving MAPF, develop suitable benchmark instances, evaluate algorithmic performance and track the state-of-the-art. The competition provides participants with a standardised system to develop, evaluate, and compare algorithmic techniques. Submissions, solutions and problem instances are all open sourced, to lower barriers, promote dissemination and enable further advancements.
Authors: Shao-Hung Chan, Zhe Chen, Teng Guo, Han Zhang, Yue Zhang, Daniel Harabor, Sven Koenig, Cathy Wu, Jingjin YuVideo
Solving the Rubik's Cube with a PDDL Planner
![](demo2.png)
Rubik's Cube (RC) is a popular puzzle that is also computationally hard to solve. In this demonstration, we introduce the first PDDL formulation for the 3x3 RC and solve it with an off-the-shelf Fast-Downward planner. Notably, we submitted this PDDL domain to the International Planning Competition (IPC) 2023's Classical track, where it emerged as one of the toughest domains to solve. We also create a plan executor and visualizer to show how the plan achieves the intended goal. Our system has two types of audiences: (a) planning researchers who can explore a hard problem and improve their planning algorithms, and (b) RC learners who want to learn how to solve the puzzle at their own pace and can now modify an initial plan (e.g., manually, using other algorithms) and see their execution.
Authors: Bharath Chandra Muppasani, Vishal Pallagani, Biplav SrivastavaVideo
Posthoc: A Visualisation Framework for Understanding Search
![](demo3.png)
We present Posthoc, a debugging and visualisation framework that helps users better understand how search algorithms work. Posthoc takes as input search traces, human-readable output logs produced by an algorithmic problem solving program. The logs are are used for subsequent playback, analysis and visualisation. Our system does not depend on any specific type of visualisation nor any particular decision-making schema. Being independent, Posthoc readily complements new and existing solvers: for AI planning, pathfinding, and heuristic search, and it can be integrated as a complementary problem-solving tool alongside.
Authors: Kevin Zheng, Daniel Harabor, Michael WybrowVideo
A Human-in-the-loop API Sequencing Tool Powered by AI Planning
![](demo4.png)
In this demo, we present a real-time API recommender system powered by an automated planner. Our tool generates multiple API recommendations for over 600 APIs within a bound of 5 seconds. By inputting a partial list of APIs, our tool dynamically fills in missing components to create a more functional or complete workflow. While this task has been historically perceived as solely a data-driven endeavor, we demonstrate how a planner can be harnessed to utilize both association information and structural dependencies between different APIs. As part of the demonstration, we also report on data gathered from the initial deployment of the tool.
Authors: Jungkoo Kang, Tathagata Chakraborti, Junkyu Lee, Michael Katz, Shirin Sohrabi, Gaodan Fang, Punleuk Oum, Prabhat Reddy, Diego del Rio, Debashish Saha, Gegi Thomas, Benjamin Herta, Jim Laredo, Alina Rotarescu, Alice Zhang, Suzette SamoojhVideo
A Demonstration of Natural Language Understanding in Embodied Planning Agents
![](demo5.png)
Autonomous agents operating in human worlds must understand and respond to natural language used by humans to communicate their tasks needs. In this paper, we present an approach for language understanding in embodied planning agents. Our approach uses recent advances in large language models to translate human language into a meaningful representation. The representation is further analyzed through grounded reasoning to connect information contained in it with the agent's current beliefs about the state of the world. Grounded reasoning results in a goal description in PDDL which is given to a planner to generate a plan and is executed. We demonstrate our approach on AI2Thor - an interactive, simulated home domain that is becoming a standard benchmark for conversational embodied agents.
Authors: Sachin Grover, Shiwali MohanVideo
Self-adaptive Mission Planning using High-Fidelity Open World Simulation
![](demo6.png)
AI and ML agents are developed with closed world assumptions, that can change during execution. This demo paper presents HYDRA, a framework for developing self-adaptive autonomous agents capable of handling unexpected domain shifts (also called \textit{novelty}) during execution, applied to a high fidelity simulator for military mission planning. The framework is divided into a base agent, responsible for basic predict-decide-act cycle, and novelty monitoring to detect, characterize and adapt to the novelty. AFSIM is a high-fidelity mission simulator that incorporates many real-world military models; and has been used for mission planning in several scenarios. This paper shows successful integration of HYDRA with AFSIM, and demonstrates HYDRA agents efficiently adapting to novelty in realistic simulated military scenarios. Demonstration of our system is available at https://tinyurl.com/wb3z2edv.
Authors: Wiktor Mateusz Piotrowski, James Chao, Sachin Grover, Roni Stern, Shiwali Mohan, Douglas S. LangeVideo
FRICODILE: Providing FRICO with Dialogue Capability
![](demo8.png)
Besides observing the pilot passively and providing contextual assistance in light of carrying out various flight tasks more efficiently, engaging the pilot in a dialogue creates an opportunity to go beyond passive monitoring of pilot's activities and flight status, by actively requesting for input information from the pilot, enriching therefore the knowledge of the pilot assistance system on the context. Such a dialogue has to be dynamic, concise and goal-directed, as it can be safety critical. In this paper, we demonstrate how FOND-planning-based dialogue agents can be applied to the use case, while also leveraging recent advances in natural language understanding and text-to-speech synthesis, in order to generate instances for the computation of plans as guidance to the pilot in an automated manner problem.
Authors: Prakash Jamakatel, Rebecca de Venezia, Christian Muise, Jane Jean KiamVideo
A Visual Studio Code Extension for Automatically Repairing Planning Domains
![](demo9.png)
We demonstrate a Visual Studio Code extension which aims at providing modeling assistance for modeling planning domains in PDDL, which serves as a front-end of our previous work. The extension can identify potential flaws in a domain and propose respective corrections by taking as input a set of counter-example plans, which are known to be valid but actually contradict the domain. Those input plans shall be provided by the user. The flaws are then identified and corrected by making changes to the domain so as to turn those plans into solutions, i.e., the changes are regarded as potential corrections to the domain. The extension supports corrections that add predicates to or remove predicates from actions' preconditions and effects.
Authors: S.T. Lin, Mohammad Yousefi, Pascal Tobias BercherVideo
Simulating Robotics Planning Domains with PDSim and ROS
![](demo10.png)
This paper describes work on the Planning Domain Simulation System (PDSim) and its application to robotics. PDSim is a plugin for the Unity game engine for visualising planning domains and plans. PDSim's original system design translates the output of a planner to 2D or 3D animations and effects. PDSim aims to assist users in evaluating the quality of a plan and improve domain and problem modelling. This system demonstration outlines the basic structure of PDSim and how to integrate it with the Robotics Operating System (ROS) to simulate plans in robotics domains. An example with a robotic arm is used to showcase how to interface with ROSbags to be able to visualise sensors from an Internet of Things (IoT) environment used for daily assistive living. Furthermore, the demo also shows how to interact with the robotics PDSim API for plan repair and replanning.
Authors: Emanuele De Pellegrin, Ron PetrickVideo1 Video2
Operations for Autonomy Tools: Task Network Editor, Plan Reconstruction Visualizer and Testbed
![](demo11.png)
In this system demonstration we present at set of tools being developed at NASA Jet Propulsion Laboratory to support the next generation of operations of autonomous spacecrafts, specifically those capable of performing onboard planning and scheduling. We will focus on demonstrating two critical tools, Task Network Editor and Plan Reconstruction, that represent key functions in the uplink and downlink processes. In the uplink process, the Task Network Editor supports the knowledge engineering process of capturing intent/goals from engineers, scientists, and operators to be sent to the spacecraft. In the downlink process, the Plan Reconstruction tool supports the understating the decisions made by the onboard planner and the execution status. These tools have been integrated to an autonomy testbed, called MASCOT, to study the interaction between ground and onboard planning, as well as to assess/characterize the planner's performance running in different processors and hardware settings.
Authors: Tiago Vaquero, Bennett Huffman, Lini Mestar, Alberto Candela, Federico Rossi, Lorraine M Fesq, Rebecca CastanoVideo
Strategic Sorcery: Automated Planning for ‘Magic: The Gathering’
![](demo12.png)
‘Magic: The Gathering’ (MTG) is a strategic card game that includes complex interactions and competitive play, using a catalogue of over 20,000 unique cards. In this demonstration, we explore how numeric planning can be used to tackle an interesting subset of the game mechanics. Strong gameplay strategies are produced using the ENHSP planner, and our work serves as both an interesting testbed for numeric planners and a foundation for more elaborate strategic card-game settings modelled in PDDL.
Authors: Nicholas Tillo, Raksha Rehal, Christian MuiseVideo