PaperlayerView exact revision on arXiv

Source-linked AI summary

Skill Sequence Planning for Collaborative Multi-Robot Construction

Xi Wang, Bo Fu, Carol C. Menassa, Vineet R. Kamat, Min Deng

arXiv:2609.25649v1cs.RO

TL;DR

Long-horizon construction tasks require interdependent skills whose sequence and robot assignment vary with the target and jobsite state. The paper generates target-specific skill plans from an updated construction graph and connects them to robot motion and human review. In a simulated wooden-frame case study, robots completed planned manipulation and nailing in both complete-plan and stepwise-approval modes.

  • Problem

    Construction skill sequences and robot assignments vary with targets, materials, jobsite state, and required operations, making manual programming for each condition a barrier to flexible deployment.

    Source: §I · p. 1

  • Method

    A central controller uses a construction relationship graph and reusable PDDL domain to generate target-specific skill sequences, assign capable robots, and connect motion skills to collision-free trajectories for human-supervised execution.

    Source: §I · p. 1, A. System Framework · p. 2, C. Symbolic Skill Sequence Planning · p. 3

  • Results

    In simulation, both complete-plan and step-by-step approval modes completed the planned manipulation and nailing sequence.

    Source: §III · p. 4

  • Takeaways & Limitations

    The wooden-frame case study demonstrates the method’s integration of evolving construction information with human-supervised execution by a heterogeneous robot team.

    Source: §IV · p. 4, §III · p. 4

    The method plans one target at a time, assumes the building model provides global assembly dependencies, and supports sequential rather than concurrent multi-robot execution.

    Source: §IV · p. 4

Abstract

from arXiv · show

Robots have significant potential to automate construction processes. However, their industry adoption remains limited, partly because of the programming effort required to adapt robots to diverse tasks. This paper presents a skill sequence planning method that enables a heterogeneous team of multi-functional robots to collaboratively perform construction assembly work using reusable, preprogrammed skills such as grasping, drilling, and fastening. A central controller transforms the digital representation of the building into a construction relationship graph that represents construction entities, their states, and their parent-child relationships. Based on this representation, the system selects the next construction target, generates a symbolic sequence of skills for capable members of the robot team, and produces collision-free geometric motion plans for skill execution. The symbolic planning problem is dynamically regenerated as the construction state changes. An interactive digital twin presents the planned skill sequence and robot states to human co-workers for review and approval before execution. The method is evaluated through a construction assembly case study. By reducing the need to program robots separately for each task variation, the proposed approach supports more flexible deployment of collaborative robot teams in construction.

I. INTRODUCTION

Construction requires long, interdependent skill sequences whose actions and assignments vary with the target and jobsite state. This paper addresses that challenge with dynamically generated plans for a heterogeneous robot team.

I. INTRODUCTION

  • The planning challenge is to choose and assign each next skill across robots with different capabilities as targets, materials, and jobsite conditions vary.Manually programming a complete sequence for every task and operating condition limits flexible deployment.Source: §I · p. 1
  • Demonstration-based methods depend on representative examples, while language-model plans require grounding and verification against execution constraints.Hybrid language-model planning also depends on the accuracy of its generated planning representation; structured symbolic planning makes preconditions and effects explicit.Source: §I · p. 1
  • The method dynamically plans from current and intended construction states, assigns skills to capable robots, and links symbolic sequences to collision-free motion execution.The planning representation is updated as components are installed, materials become accessible, or the actual jobsite state deviates from expectations.Source: §I · p. 1
  • A simulation-based wooden-frame assembly case study uses two robots with complementary manipulation and fastening capabilities.The study involves two KUKA KR 120 manipulators, as described in the case-study introduction.Source: §I · pp. 1–2

A. System Framework

The system connects building information to collaborative robot execution through target selection, symbolic skill sequencing, and geometric motion planning. A human-facing digital twin supports review and approval.

A. System Framework

  • Three planning levels select the next workpiece, determine an ordered sequence of robot skills, and convert motion-related skills into collision-free trajectories.The sequence planner identifies robot types capable of performing the required skills.Source: A. System Framework · p. 2
  • The digital building model stores as-designed and as-built information in Target, Material, As-Built, As-Designed, and Virtual Collision layers.It includes geometry and attributes such as construction order, pose, required process, operation method, and construction parent; the representation is updated as construction progresses.Source: A. System Framework · p. 2
  • The planning digital twin uses MoveIt for collision-free robot motions, while the Unity interaction twin visualizes robot and construction states and presents the skill sequence.The central controller communicates with other modules through ROS.Source: A. System Framework · p. 2
  • Workers can approve the complete sequence or require confirmation before individual actions, creating a human-supervised workflow.Figure 1 is identified as the system framework.Source: A. System Framework · p. 2, Figure 1 · p. 2

B. Construction State Representation

The construction relationship graph represents entities, their states, and parent-child relationships so the controller can select accessible materials and update planning information as assembly proceeds.

B. Construction State Representation

  • The controller plans one queued target at a time, using construction-order attributes while retaining other components as context for state representation and motion planning.After installation, the selected workpiece leaves the active queue and the next target becomes available.Source: B. Construction State Representation · p. 2
  • Graph nodes represent construction entities with geometry, pose, type, layer, and state attributes, while directed edges represent construction parent-child relationships.A Jobsite node is the default root when a component has no specified construction parent.Source: B. Construction State Representation · p. 2, Figure 2 · p. 2
  • For a target requiring a material type, the controller selects a matching accessible leaf from a stack; if several match, it chooses the one nearest the responsible robot.Treating the piece below as the construction parent makes the top piece the accessible leaf.Source: B. Construction State Representation · pp. 2–3
  • After installation, the graph changes the material node to As-Built, assigns its construction parent, and exposes the next piece in its original stack.The updated graph supplies the objects, attributes, and relationships for the next symbolic planning problem; Figure 2 depicts graph structure and update.Source: Figure 2 · p. 2, B. Construction State Representation · p. 3

C. Symbolic Skill Sequence Planning

The controller combines a reusable PDDL domain with automatically regenerated, target-specific problems to plan parameterized robot skills. The resulting actions are translated into robot commands and collision-free trajectories, with human review and replanning supported.

C. Symbolic Skill Sequence Planning

  • The reusable PDDL domain defines robot types, predicates, and actions, while an automatically generated problem describes the current target and jobsite state.The problem is regenerated for each target and when relevant state changes; the domain remains unchanged across targets.Source: C. Symbolic Skill Sequence Planning · p. 3
  • Robot subtypes in action parameters and preconditions restrict each symbolic skill to team members with the required capabilities.Parameterized actions map to executable ROS skills, while symbolic places are resolved to 6DOF poses when preparing motion planning.Source: C. Symbolic Skill Sequence Planning · p. 3
  • The controller grounds domain predicates against the graph to generate PDDL objects, initial state, and goal state for the active target.The problem file is regenerated when the active target or observed construction state changes.Source: C. Symbolic Skill Sequence Planning · p. 3
  • Fast Downward produces parameterized actions assigned to capable robots; processors convert them into ROS commands and natural-language descriptions for review.Motion-related skills are resolved into poses and collision-free trajectories; if motion planning fails, the issue is escalated to the human co-worker.Source: C. Symbolic Skill Sequence Planning · p. 3
  • A worker can choose a different feasible action, after which the observed state and planning problem are updated and the skill sequence regenerated.After target completion, the digital model and relationship graph are updated before planning advances to the next target.Source: C. Symbolic Skill Sequence Planning · p. 3

III. CASE STUDY

A simulation-based case study tested the end-to-end planning and human-supervised execution of a wooden-frame assembly by a heterogeneous robot team. The robots completed the planned manipulation and nailing sequence with either complete-plan or step-by-step approval.

III. CASE STUDY

  • Two KUKA KR 120 manipulators used different end-effectors: one handled components, and the other performed connection or finishing operations.The second robot used exchangeable tools for nailing, screwing, and painting.Source: §III · p. 3
  • The wooden frame used two supporting studs and three transverse studs, each secured across the supports with four nails.The environment was simulated in Gazebo, while a Unity interface presented the evolving construction state and planned skills for human review.Source: §III · p. 4
  • In simulation, the robots completed the planned manipulation and nailing sequence with both complete-plan execution and step-by-step approval.Robot and construction states were reflected in the interaction digital twin.Source: §III · p. 4

IV. CONCLUSIONS

The method connects evolving construction information to human-supervised execution by heterogeneous robots, and a wooden-frame simulation demonstrated the integrated workflow. Its current scope is limited to one target at a time and sequential execution, with physical validation and scalability evaluation left for future work.

IV. CONCLUSIONS

  • In simulation, two robots with complementary capabilities completed manipulation and nailing under both complete-plan and step-by-step approval modes.The case study integrated target selection, symbolic planning, robot assignment, geometric motion planning, and human approval.Source: §IV · p. 4
  • A construction relationship graph and reusable domain model support target-specific PDDL generation, skill sequencing, robot assignment, and collision-free trajectories.The graph represents construction entities, their states, material accessibility, and parent-child relationships.Source: §IV · p. 4
  • The method plans for one target at a time, assumes the building model supplies global dependencies, and supports sequential rather than concurrent execution.Future work includes quantitative feasibility evidence, physical-robot validation, and evaluation with larger teams and more diverse construction processes.Source: §IV · p. 4
Loading 2609.25649v1…