On-the-move mobile grasping
Centre for Systems and Control, IIT Bombay · May 2026 – Jul 2026
First-author manuscript in preparation
Abstract
Grasping while the base is still moving removes the stop-and-settle that dominates the cycle time of most mobile manipulators, but it makes the target pose a moving quantity. The published approach predicts its correction once, before motion, and executes it open-loop. This work replaces that with a controller that closes the loop on measured base pose, solving arm and base together as a single quadratic program.
The problem
A mobile manipulator that must stop, settle, and only then reach spends most of its cycle waiting for the base to be still. Grasping on the move removes that wait. The cost is that by the time the arm arrives where the object was, the base has carried the gripper somewhere else.
Why the published pipeline breaks
The existing approach predicts a residual velocity correction once, before the motion begins, and then executes it open-loop. That assumption survives in simulation, where the base follows its commanded trajectory almost exactly.
On hardware it does not. The real base drifts from its command, and the correction it was handed no longer matches the situation it is applied to. The reported gap is a 14% real-world success drop against 1% in simulation at 0.20 m/s, and the controller is never told that its own assumption has gone stale.
Holistic QP formulation
Arm and base are solved together as a single quadratic program rather than as two coordinated controllers, so the redundancy of the combined system is available to the solver instead of being partitioned away by a hierarchy fixed in advance.
The reference formulation is validated on a Panda model through a standalone experiment, and factored into an importable Python module built on Robotics Toolbox with qpsolvers and OSQP. A ROS package wraps that solver for the lab machine, with the grasp predictor and base-pose estimator as separate nodes.
Where it stands
The reference QP is validated and the controller module is drafted. The full holistic formulation with manipulability and joint dampers, the 4-DOF adaptation to the OpenManipulator-X, and the TurtleBot 3 bring-up are in progress. The predictor and base-pose estimation nodes are placeholders, and hardware experiments have not started.