Multi-Arm Asynchronous Motion Planning for Manipulation Tasks

TL;DR

We plan coordinated motions for multi-arm robots that operate asynchronously: each arm can replan while it is still moving — individually or as a group — instead of waiting for the others. To produce high-quality motions in the high-dimensional configuration spaces these systems live in, we pair sampling-based planning with trajectory optimization over Graphs of Convex Sets (GCS) that are built from prior planning experience and continuously refined as the system plans more. On a dual-arm food-plating system, this yields substantially higher-quality coordinated trajectories than classical sampling-based planners, and running the arms asynchronously significantly increases throughput by 34.47%.

Abstract

Multi-arm robotic systems enable coordinated manipulation and parallel task execution, offering substantial gains in capability and throughput relative to single-arm robots. Realizing these benefits, however, requires motion planning in high-dimensional configuration spaces while coordinating asynchronous and heterogeneous robot behaviors. We present a whole-system approach that captures both local and global motion requirements, supports coordinating multiple arms asynchronously, and enables all or a subset of arms to replan while in motion. To generate high-quality motions in high-dimensional spaces, we combine sampling-based methods with trajectory optimization over Graphs of Convex Sets (GCS) that are constructed from prior planning experience and are continuously refined to improve future planning performance. We demonstrate the approach on a dual-arm food-plating platform that combines expressive learned manipulation skills with fast, reliable transit motions. Experiments show that the proposed GCS-based planner generates substantially higher-quality coordinated trajectories than classical sampling-based methods and that asynchronous operation significantly improves throughput compared to sequentially moving arms.

The Problem

Multiple arms sharing a workspace can manipulate cooperatively and execute tasks in parallel — more capability, more throughput. The cost of those benefits is the planning problem underneath:

• High-dimensional configuration spaces. Every added arm enlarges the joint space the planner must search, and coordinated tasks couple the arms so they cannot simply be planned in isolation.
• Asynchronous, heterogeneous behaviors. Arms run different skills on different timelines. A planner that forces them onto a single shared clock wastes time waiting on the slowest arm; letting them run freely instead means each arm is a moving obstacle to the others and must plan against information that is constantly going stale.
• Replanning in motion. Useful systems need to adapt mid-task — ideally letting any arm, or a subset, replan without stopping the whole cell.

Approach

We take a whole-system view of planning: rather than treating each arm in isolation, the formulation captures both local motion requirements (each arm's own constraints and skills) and global ones (how the arms must coordinate). On top of that, the planner supports asynchronous coordination and lets all — or just a subset of — arms replan while in motion.

The core of the method generates high-quality motions in high-dimensional spaces by combining two ingredients:

1. Sampling-based exploration to discover feasible motion in the high-dimensional configuration space.
2. Trajectory optimization over Graphs of Convex Sets (GCS). The GCS is constructed from prior planning experience and continuously refined as the system plans, so planning quality and speed improve over time rather than starting from scratch on every query.
3. Asynchronous coordination. Arms commit to and execute motions on their own timelines, with replanning available to any arm or subset mid-task — which is what turns the planner's quality into end-to-end throughput.

Key insight. Reusing and continuously refining planning experience as a Graph of Convex Sets lets us optimize high-quality coordinated trajectories in spaces where sampling alone struggles — and decoupling the arms asynchronously converts that trajectory quality into real throughput gains.

The Platform: Dual-Arm Food Plating

We demonstrate the approach on a dual-arm food-plating platform that combines expressive learned manipulation skills (the dexterous plating motions) with fast, reliable transit motions (moving between stations). We present a dual-arm robotic system with the ultimate target of plating a Michelin-three-star quality dish. The execution speed is critical as the food quality is highly time-sensitive. For example, cooked salmon needs to be plated and served within minutes to preserve its temperature, taste and texture quality. Additionally, some ingredients require specialized tools that must be picked up from a holder before use and returned afterward. The plate, tool holders, and nearby manipulation zones form a densely shared workspace in which independent runs from arms can collide with one another, requiring a sophisticated motion planner that can replan quickly while avoiding collision.

Experiments & Results

Quantitative results of making the River Plate

Above Table Sync vs. async timing in simulation-only benchmarking (no real hardware or vision; task planning and motion time only), comparing planner configurations with GCS planning enabled (on) versus solely sample-based planning (GCS off). The GCS-on run used all successful trials (45 paired comparisons). The GCS-off planning run reflects the updated filtered benchmark set (40 paired comparisons). Overall speedup is computed as 1 − μasync/μsync.

Qualitative results

Synchronous vs. asynchronous operation (plating the diced veggies), side by side:

Citation

The paper is not published yet. Until it is, a preprint citation (this is an indented code block — keep the 4-space indent so it renders without a fenced-code extension):

@unpublished{yourkey2026multiarm,
  title  = {Multi-Arm Asynchronous Motion Planning for Manipulation Tasks},
  author = {Sadra Sadraddini, Ramy Yammine, David Bambrick, Robert Truax, Jackie Videira, Alan Zhou, Bob El-Masri, Brian Krieger, Rana Odabas, Brook Stevens, and Peter Stone},
  note   = {Preprint. Under review},
  year   = {2026}
}

When it appears, switch to the published @inproceedings{...} form with the booktitle and year.