Centralized vs. Decoupled Dual-Arm Planning Taking into Account Path Quality

TL;DR

The paper analyzes centralized versus decoupled dual-arm planning and finds that, for standard pipelines like MoveIt!, centralized planning generally outperforms decoupled approaches due to lower computation time and higher robustness. It then introduces a rotation-aware path-length optimization by extending PLPP to SO(3) with analytic gradients, and demonstrates that integrating PLPP into a decoupled framework significantly improves path quality and robustness, achieving up to 99.9% success with parallel planning pipelines. The work also provides a detailed benchmark on a real dual-arm mobile system, supported by SSV-based distance computations and hardware experiments, highlighting online viability. Overall, the study clarifies planning strategy trade-offs for dual-arm manipulation and offers a practical method (PLPP with rotational support) to enhance path quality while maintaining online performance.

Abstract

The aim of coordinated planning is to avoid robot-to-robot collisions in a multi-robot system, and there are two standard solution approaches: centralized planning and decoupled planning. Our first contribution is a decoupled planning approach that ensures C2-continuous control commands with zero velocities at the start and goal. We benchmark our decoupled approach with a centralized approach. Contrary to literature, we show that for a standard motion planning pipeline, such as the one used by MoveIt!, centralized planning is superior to decoupled planning in dual-arm manipulation: It has a lower computation time and a higher robustness. Our second contribution is an optimization that minimizes the rotational motion of an end-effector while considering obstacle avoidance. We derive the analytic gradients of this optimization problem, making the algorithm suitable for online motion planning. Our optimization extends an existing path quality improvement method. Integrating it into our decoupled approach overcomes its shortcomings and provides a motion planning pipeline that is robust at up to 99.9% with a planning time of less than 1s and that computes high-quality paths.

Figures (11)

• Figure 1: The hardware of our mobile dual-arm system integrates three components: a custom-built mobile base, two Franka Emika robots, and the robominds vision system. The camera is mounted on the left end-effector.
• Figure 2: Step \ref{['enum:PickAssembly']} of the pick-and-place evaluation scenario defined in Section \ref{['sec:PickAndPlace']}. We evaluate on the real robot (Figure \ref{['fig:real']}) and in simulation (Figure \ref{['fig:sim']}) and we use our SSV-based environment modeling approach (Figure \ref{['fig:ssv']}). The default FCL-based environment modeling approach of MoveIt! is shown for comparison (Figure \ref{['fig:fcl']}).
• Figure 3: Planning pipeline of our centralized approach with the four main modules shown in light blue: Path Planner, Path Post-Processor, Trajectory Planner, and Motion Controller. The planning pipeline receives motion requests via ROS and computes joint velocity commands for the two manipulators.
• Figure 4: Planning pipeline of our decoupled approach with the four main modules shown in light blue: Path Planner, Path Post-Processor, Trajectory Planner, and Motion Controller. There are two instances of the first three: one for each arm. There is one Motion Controller that coordinates the two calculated trajectories. Like the centralized approach in Figure \ref{['fig:CentrPipeline']}, the planning pipeline receives motion requests via ROS and computes joint velocity commands for the two manipulators.
• Figure 5: Three different interpolation approaches for the coordination space trajectory $\mathbf{c}(t)$. The black dashed vertical lines show the path waypoints of the fixed-path coordination. That is, these are the coordination time stamps of the right robot computed by the RRTconnect and path simplifier in $\mathcal{C}$.