PaperlayerView exact revision on arXiv

Source-linked AI summary

cuArms: GPU-Accelerated Conflict-Based Planning for Multi-Arm Manipulation

Philip Huang, Fabio Ramos, Jiaoyang Li, Caelan Garrett, Lars Johannsmeier

arXiv:2610.36250v1cs.ROcs.AIcs.MA

TL;DR

Multi-arm planning must coordinate task assignments, schedules, motions, and manipulation skills, while prior approaches can be slow, domain-specific, or poorly scalable. cuArms combines GPU-accelerated conflict-based search with reusable per-arm roadmaps and a high-level scheduler. Across welding and tabletop tasks, it finds initial plans in real time and improves makespans compared with the reported baselines.

  • Problem

    Existing multi-arm approaches can be too slow, depend on domain-specific assumptions, or scale poorly with robot count and task horizon.

    Source: §I · p. 1

  • Method

    cuArms uses GPU-batched conflict-based search over reusable per-arm multi-modal roadmaps and grounds scheduler-proposed task skeletons into coordinated plans.

    Source: §I · p. 1, §V · p. 6

  • Results

    In tabletop TAMP, cuArms finds initial solutions 46.4% faster than ScheduleStream and reduces median makespans by 42–57% across four tasks.

    Source: B. Multi-Arm TAMP · p. 7

  • Takeaways & Limitations

    Near-instantaneous solutions let users rapidly evaluate multi-arm task and workspace configurations and integrate domain-specific costs and collaborative skills.

    Source: §VII · p. 8

Abstract

from arXiv · show

Multi-arm manipulation has the potential to dramatically improve the throughput, efficiency, and dexterity of many manipulation tasks. Recent work coordinates multiple robot arms in a shared workcell by allocating and scheduling tasks for greater parallelism, jointly planning multi-arm motions, or directly learning multi-arm policies. So far, these approaches are either too slow, rely on domain-specific assumptions, or scale poorly with the number of robots and task horizon. Thus, we introduce cuArms, a conflict-based planner co-designed with GPU parallelism that efficiently decomposes the search space using reusable multi-modal roadmaps for individual robots and resolves conflicts and deadlocks through a lazy search. To jointly optimize task assignment, schedules, motions, and manipulation skills, we integrate a high-level scheduler that proposes candidate skeletons for cuArms to ground into coordinated manipulation plans. Experimental results show that our method establishes a new state of the art, with real-time initial solutions and anytime improvements in both motion quality and scheduling, without days of domain-specific training required by learning approaches. Videos of both simulated and real-world industrial tasks are available on https://cuarms.github.io/.

I. INTRODUCTION

Multi-arm manipulation can improve throughput, space utilization, and dexterity, but coordinating task allocation, schedules, collision safety, and hybrid skills remains challenging. cuArms combines GPU acceleration with conflict-based planning and integrates with a high-level scheduler to coordinate multi-arm tasks.

  • Multiple arms can increase task throughput and factory-floor space utilization, and support dexterous coordination or bulky-part handling.Source: §I · p. 1
  • Existing approaches are often domain-specific, time-consuming, or unable to plan bimanual skills, while multi-arm coordination must address allocation, scheduling, safety, and hybrid action spaces.Source: §I · p. 1
  • cuArms combines GPU acceleration with conflict-based multi-agent planning and reusable per-arm roadmaps to search long-horizon, multi-modal, multi-arm plans.Source: §I · p. 1
  • Initial plans take about 1 second across varied domains, including an 8-arm, 40-task welding task, and additional computation improves solution quality by up to 57%.The experiments include simulation and real-robot tasks.Source: Figure 1 · p. 1, §I · p. 1
  • cuArms integrates with a high-level scheduler to support domain-agnostic task and motion planning, including collaborative bimanual skills such as handover and carry.Source: §I · p. 1

II. RELATED WORK

Prior work spans multi-agent path finding, multi-arm manipulation planning, and hardware-accelerated planning. cuArms adapts lazy conflict-driven search to contact-rich arm tasks and batches GPU collision checks to address computational bottlenecks.

A. Multi-Agent Path Finding (MAPF)

  • LaCAM embeds PIBT conflict resolution in a complete lazy joint-space search, while cuArms applies LaCAM-style search to contact-rich multi-arm manipulation with GPU-accelerated search and roadmap construction.MAPF methods coordinate robots on discrete space-time graphs; LaCAM-style planning had previously been shown for mobile robots.Source: A. Multi-Agent Path Finding (MAPF) · p. 2

B. Multi-Arm Manipulation Planning

  • Prior multi-arm planners address single-query motions or long-horizon pick-and-place and handover, while cuArms jointly searches coupled roadmaps for asynchronous trajectories and skills without task-specific hierarchies.Collision checking is a dominant bottleneck for several prior multi-arm planning approaches.Source: B. Multi-Arm Manipulation Planning · p. 2

C. Accelerated Motion Planning

  • Unlike CPU-vectorized multi-arm collision checking, cuArms evaluates collisions early in a few large GPU batches to reduce the bottleneck while avoiding excessive CPU-GPU communication.VAMP-MR vectorizes collision checking with CPU AVX2 instructions, but collision checking remains its computational bottleneck.Source: C. Accelerated Motion Planning · p. 2

III. MULTI-ARM MANIPULATION PLANNING

The planning formulation represents the workspace through robot configurations, object modes, and manipulation skills. A schedule skeleton specifies actions and their ordering, while cuArms searches for coordinated robot movements and skills.

  • The workspace contains robot arms, static obstacles, and movable rigid objects; each arm has a configuration space subject to joint-position and velocity limits.Source: §III · p. 2
  • A mode specifies each object's grasp, attachment, or placement relative to a robot or the world, and manipulation skills can change modes.Source: §III · p. 2
  • Manipulation skills are parameterized, contact-rich interactions involving one or two arms, with entry and exit configurations, minimum duration, collision bounds, and a mode-transition function.Source: §III · p. 2
  • A schedule skeleton fixes the actions, participating robots, skill parameters, and precedence constraints; actions on different robots may execute asynchronously when unordered.Source: §III · p. 2
  • The system overview has a scheduler propose task assignments, ordering, and satisfying samples, after which cuArms builds per-robot multi-modal roadmaps and searches for conflict-free movements and skills.The overview also shows trajectory refinement by shortcutting and replanning with different skeletons.Source: Figure 2 · p. 3, §III · p. 3

A. Problem Formulation

Given a workspace and schedule skeleton, the problem is to jointly plan collision-free move trajectories and schedule actions while respecting the skeleton’s constraints. A feasible timed plan must also satisfy collision, precedence, duration, and waiting requirements.

A. Problem Formulation

  • A solution jointly schedules actions and plans move trajectories that connect skill configurations while respecting the mode sequence induced by skills.The schedule skeleton fixes actions, participating robots, skill parameters, and precedence constraints.Source: A. Problem Formulation · p. 3
  • A plan is feasible only if moving, waiting, and active-skill geometries avoid obstacles and other robots, and skill durations meet their minimums.Precedence requires each earlier action to finish before the next begins; robots may wait while maintaining attachments and collision constraints.Source: A. Problem Formulation · p. 3
  • The objective combines makespan, the sum of robot completion times, and motion-distance costs to support domain-specific preferences.Distance may be measured in joint space, task space, collision-sphere-weighted distance, or a weighted combination.Source: A. Problem Formulation · p. 3

IV. CUARMS ALGORITHM

cuArms separates coordination into roadmap construction, joint coordination, and anytime refinement, with each phase designed around GPU-batched geometric computation. Its per-robot multi-modal roadmaps support motion and skills while reusing geometry across modes and schedule skeletons.

IV. CUARMS ALGORITHM

  • cuArms separates planning into roadmap construction, joint coordination, and anytime refinement, co-designing the phases around GPU-batched geometry and conflict-based coordination of skills and motions.This differs from independently generated or sequentially generated schedules and composite-space sampling with prioritized planning,,.Source: §IV · p. 3

A. Multi-Modal Roadmap Construction

  • Per-robot probabilistic roadmaps include free-space motions and skill-specific edges, with valid nodes and edges sampled and stored in GPU memory.Cached geometry can be reused across schedule skeletons; mode-validity masks check object interactions without rebuilding roadmaps.Source: A. Multi-Modal Roadmap Construction · pp. 3–4
  • Mode-validity masks reuse sampled roadmaps across modes by checking cached configurations against mode-specific object placements and attached-object interactions in GPU batches.The decomposition reuses forward kinematics and robot-environment collision results; the stated rationale is that objects sparsely affect roadmap validity.Source: A. Multi-Modal Roadmap Construction · p. 4
  • Skill entry configurations become roadmap nodes and skill edges encode execution and mode transitions, connecting neighboring moves through entry and exit configurations.Skill bounding volumes support batched collision checks, and durations are rounded to multiples of the maximum edge interval for synchronization.Source: A. Multi-Modal Roadmap Construction · p. 4
  • Cost-to-go heuristics guide search from each roadmap node and current action toward its target, while a separate cost-to-home heuristic can guide blocked robots toward a conflict-free home configuration.Backward Dijkstra uses unit edge costs on the base roadmap without the mode-validity mask.Source: A. Multi-Modal Roadmap Construction · p. 4

B. GPU-Accelerated Conflict-Driven Search

cuArms extends LaCAM to track task progress and resolve multi-arm conflicts during long-horizon, multi-modal planning. It batches geometric conflict checks on the GPU before PIBT performs lazy joint-successor selection.

B. GPU-Accelerated Conflict-Driven Search

  • cuArms augments LaCAM’s lazy joint-state search with task-level progress because the original method does not track long-horizon manipulation skills.The original LaCAM uses depth-first search guided by PIBT; motion planning independently cannot resolve the illustrated deadlock by retiming alone, while sequential planning must order actions in the skeleton.Source: B. GPU-Accelerated Conflict-Driven… · p. 4
  • The search stores roadmap and action states plus action constraints, lazily exploring alternative joint successors without enumerating all action combinations upfront.Candidate motions, waits, and skills are ranked with manipulation skills first; participating robots must reach skill entry nodes and satisfy precedence constraints.Source: B. GPU-Accelerated Conflict-Driven… · pp. 4–5
  • cuArms batch-checks inter-robot candidate pairs on the GPU, then PIBT uses the Boolean conflict table during priority inheritance and backtracking.Aligned trajectories are used for motions and enclosing volumes for skills; separating recursive CPU control flow from batched GPU computation avoids repeated launches and synchronization.Source: B. GPU-Accelerated Conflict-Driven… · p. 5
  • For bimanual skills, cuArms constrains participating robots’ actions together and can prioritize yield motions when skeleton precedence blocks a move.Search ends when all actions are complete, and the path is converted into a timed plan.Source: B. GPU-Accelerated Conflict-Driven… · p. 5

C. Anytime Refinement

cuArms refines initial plans with batched multi-arm shortcuts because coarse roadmaps and LaCAM-style search lack an optimality guarantee. It checks candidate shortcuts for static and inter-robot collisions, then greedily applies non-overlapping improvements.

C. Anytime Refinement

  • Initial solutions may be suboptimal because roadmaps are coarse and LaCAM-style search lacks an optimality guarantee, motivating shortcut refinement.Multi-arm shortcuts are nontrivial for long-horizon plans because optimizing one arm must not create collisions with other robots’ motions or skills.Source: C. Anytime Refinement · p. 5
  • cuArms enumerates bounded straight-line joint-space shortcuts for each move and batch-checks them against the static environment and movable objects in the current mode.It samples a large batch of multi-arm shortcuts from valid single-robot shortcuts, including single-arm, prioritized, or concurrent composite shortcuts.Source: C. Anytime Refinement · p. 5
  • After checking sampled multi-arm shortcuts for inter-robot conflicts in parallel, cuArms greedily combines non-overlapping shortcuts in descending cost-improvement order.The algorithm returns a plan through batched refinement after finding an initial solution.Source: C. Anytime Refinement · p. 5

V. MULTI-ARM TASK AND MOTION PLANNING

The full multi-arm TAMP problem jointly chooses task assignments, schedules, skill parameters, and coordinated motions. cuArms grounds candidate schedule skeletons in coordinated trajectories, then can return to the scheduler for alternatives when needed.

V. MULTI-ARM TASK AND MOTION PLANNING

  • Jointly optimizing schedule skeletons and coordinated manipulation plans can improve on using either the high-level scheduler or manipulation planner alone.A scheduler can handle allocation, scheduling, and manipulation modes but is less effective at tightly coordinated motions; fixed-skeleton motion optimization cannot repair poor task assignments or skill parameters that cause unnecessary waiting.Source: §V · pp. 5–6
  • Fig. 5 shows execution makespan improving as cuArms shortcuts initial trajectories and searches alternative schedules.The search and refinement are applied to RoboBallet schedules.Source: Figure 5 · p. 6
  • A task-level scheduler proposes candidate skeletons, and cuArms grounds them in coordinated trajectories that support asynchronous execution.If no candidate is feasible, cuArms backtracks to request more skeletons; batches of candidates can reduce backtracking.Source: §V · p. 6
  • When candidate skeletons share manipulation modes, cuArms reuses the GPU-resident multi-modal roadmap and updates mode-validity masks for newly introduced modes.New modes do not require modifying the base roadmap.Source: §V · p. 6

VI. EXPERIMENTS

The experiments test initial-solution speed, anytime quality improvement, and scaling across simulated and real-world tasks. In RoboBallet welding, cuArms finds real-time solutions and improves makespans relative to the retrained RL policy on episodes both methods solve.

VI. EXPERIMENTS

  • The evaluation compares initial-solution time and initial and final makespan, using 15 random seeds per simulation setting unless stated otherwise.Plots report median and interquartile range; the workstation uses an NVIDIA RTX Pro 6000 GPU.Source: §VI · p. 6

A. RoboBallet: Many-Arm and Long-Horizon Manipulation

  • The RoboBallet evaluation covers 4 or 8 arms performing 10 or 40 spot-welding tasks in a tight, obstacle-filled workspace.RoboBallet is retrained from scratch in about two days; its policy has no completeness or formal safety guarantees and assumes goal tolerances of up to 2.5 cm and 15°.Source: A. RoboBallet: Many-Arm and… · p. 6, §VI · p. 6
  • 20.1% and 27.9% lower median makespans than RL on jointly solved 10-target and 40-target episodes, respectively.cuArms finds initial solutions in 0.2 s median for 10 targets and 0.41 s for 40 targets; shortcutting and alternative schedules reduce initial makespan by about 47% in both settings.Source: A. RoboBallet: Many-Arm and… · pp. 6–7, Table II · p. 6
  • When doubling to 8 arms, cuArms reduces median makespan by 10.2% while still finding feasible solutions within 1 second.GPU collision checking accounts for 18.1–24.7% of time-to-initial solution with 4 arms and 50.2% with 8 arms; connecting weld targets to the base roadmap is another major bottleneck.Source: A. RoboBallet: Many-Arm and… · p. 7

B. Multi-Arm TAMP

The evaluation tests cuArms on two-arm tabletop manipulation and finds faster initial solutions than ScheduleStream, with better makespan except for Hold. It also reports improved makespans through refinement and better scaling to larger packing tasks.

B. Multi-Arm TAMP

  • The evaluation covers holding, stacking, packing, and unpacking 2–4 cubes with two Panda arms, plus packing and unpacking with 6 and 8 objects.Both methods use ScheduleStream’s symbolic planner and a 60 s high-level search budget.Source: B. Multi-Arm TAMP · p. 7
  • For packing, cuArms scales to 8 objects, while ScheduleStream finds fewer feasible solutions and higher-makespan plans, and fails to solve 8-object packing.The scaling comparison is shown in Fig. 7.Source: Figure 7 · p. 7, B. Multi-Arm TAMP · p. 7
  • Unlike ScheduleStream’s independent arm-path planning and conflict-avoidance retiming, cuArms jointly coordinates manipulation skills and motion trajectories across candidate schedule skeletons.It refines initial trajectories for 2 s per packing skeleton and 1 s per skeleton in other tasks, optimizing weighted makespan and path-distance costs.Source: B. Multi-Arm TAMP · p. 7
  • Initial solutions take 0.48–2.96 s, 46.4% faster than ScheduleStream, with better initial makespan except for Hold; cuArms solves all instances while ScheduleStream sometimes fails.Multi-arm shortcutting and alternative-candidate exploration reduce median cuArms makespan by 42–57% across the four tasks; Table IV reports initial and final makespan as J0 and Jf.Source: B. Multi-Arm TAMP · p. 7, Table IV · p. 7

C. Real-Robot Electronics Part Assembly

The real-robot evaluation applies cuArms to two electronics-assembly tasks with two Flexiv Rizon4 arms. The conclusion presents cuArms as a conflict-based planner that rapidly finds solutions and can support feasibility evaluation across task and workspace configurations.

C. Real-Robot Electronics Part Assembly

  • The electronics-assembly system uses GPU-accelerated skill sampling for constrained poses and a ScheduleStream-based symbolic planner to generate schedule skeletons.cuArms produces coordinated plans with collision-free approach motions around camera mounts; the tasks use two Flexiv Rizon4 arms.Source: C. Real-Robot Electronics Part Assembly · pp. 7–8

VII. CONCLUSION

  • cuArms combines reusable per-arm roadmaps, inter-robot conflict resolution, and GPU-batched geometry evaluation in a conflict-based multi-arm planner.The conclusion describes integrating it into a schedule-then-coordinate TAMP system for low-level coordination and high-level scheduling.Source: §VII · p. 8
  • The conclusion identifies rapid feasibility evaluation across multi-arm task and workspace configurations as a use for near-instantaneous solutions.It also notes integration with domain-specific costs and collaborative skills.Source: §VII · p. 8
Loading 2610.36250v1…