PaperlayerView exact revision on arXiv

Source-linked AI summary

QuadHand: A Compact Quadrotor Aerial Manipulator with MRC-SDF-Based Whole-Body Motion Planning

Rui Jin, Ruiyang Liu, Xinhang Xu, Haotian Jin, Yi Wang, Yizhuo Yang, Lihua Xie

arXiv:2609.35094v1cs.RO

TL;DR

Aerial manipulators must manage arm-induced disturbances while retaining geometric detail for close-proximity planning and coordinating a high-DoF base–arm system. QuadHand combines passive CoG compensation, MRC-SDF, and coupled whole-body trajectory optimization; MRC-SDF reached 99% success compared with 0% for ESDF in the reported benchmark.

  • Problem

    Close-range aerial manipulation requires managing arm-induced disturbances, preserving geometry around targets, and planning the coupled high-DoF base and arm.

    Source: §I · p. 1

  • Method

    QuadHand combines a 3-DoF arm and gripper with battery-assisted passive CoG compensation, MRC-SDF geometry, and coupled spatiotemporal whole-body trajectory optimization.

    Source: §I · pp. 1–2, §VIII · p. 7

  • Results

    99% success rate for MRC-SDF versus 0% for ESDF, with 172.22 ms average optimization time under successful MRC-SDF trials.

    Source: C. Simulation and Benchmark Comparisons · p. 6

  • Takeaways & Limitations

    Experiments demonstrate safe aerial manipulation in complex environments using the proposed compact platform and whole-body planner.

    Source: §VIII · p. 7

    The system relies on external motion capture and manually predefined terminal states and retreat directions.

    Source: §VIII · p. 7

Abstract

from arXiv · show

Uncrewed aerial manipulators (UAMs) integrate robotic arms with aerial platforms for three-dimensional physical interaction. However, enlarging the workspace increases arm-induced disturbances, while existing geometric representations face a trade-off between geometric fidelity and computational efficiency in close-proximity interaction. This paper presents QuadHand, a compact quadrotor aerial manipulator with a 3-DoF arm, gripper, and battery-assisted passive CoG compensation module to reduce dominant arm-induced disturbances. We further propose MRC-SDF, a Multi-articulated Robot-Centric Signed Distance Field that preserves fine geometric detail with tractable computation, and a spatiotemporal whole-body trajectory optimization framework that jointly optimizes the quadrotor and manipulator for safe and executable trajectory generation. Simulations and real-world experiments demonstrate safe and executable aerial manipulation in complex environments.

I. INTRODUCTION

Aerial manipulators enable physical interaction in three-dimensional workspaces, but face coupled challenges in controlling arm-induced disturbances, preserving useful geometry near targets, and planning the base and arm together. QuadHand combines a compact compensated platform, MRC-SDF geometry, and coupled whole-body optimization, validated in simulation and real-world flight tests.

I. INTRODUCTION

  • As the end-effector workspace grows, arm motion and interaction forces can increase aerial-base disturbances, making workspace and control accuracy a design trade-off.The introduction identifies balancing workspace against base disturbance as crucial for end-effector control accuracy.Source: §I · p. 1
  • Close-proximity interaction requires geometric representations that preserve robot and environment detail for safe, precise operation in complex environments.The introduction notes that geometric information loss constrains the feasible solution space.Source: §I · p. 1
  • Because the coupled base and arm have high DoFs, planning them independently can restrict the solution space and yield suboptimal or infeasible trajectories.Source: §I · p. 1
  • QuadHand combines a quadrotor, 3-DoF arm, gripper, and battery-assisted passive CoG compensation with MRC-SDF geometry and whole-body trajectory optimization.The paper reports simulations and real-world flight tests demonstrating system efficiency.Source: §I · pp. 1–2

II. RELATED WORKS

The paper situates QuadHand among aerial platforms and whole-body planning representations, then describes its architecture and hardware. Its design pairs a 3-DoF serial arm with passive CoG compensation and planning that retains environmental point-cloud geometry.

A. Quadrotor Aerial Manipulation Platform

  • Serial arms offer greater workspace and dexterity than parallel manipulators, but bring longer moment arms, higher joint torques, and larger CoG shifts; QuadHand uses battery-assisted passive compensation.Parallel manipulators offer higher positioning accuracy but limited workspace.Source: A. Quadrotor Aerial Manipulation… · p. 2

B. Geometric Representations for Whole-Body Planning

  • Corridor and convex-region abstractions can lose robot or target-object geometry, while fine-resolution ESDFs preserve detail at higher memory and construction cost.These trade-offs matter for close-range interactions where contact clearance and local geometry are critical.Source: B. Geometric Representations for… · p. 2

III. SYSTEM ARCHITECTURE

  • The system estimates quadrotor state by fusing motion-capture pose measurements with onboard IMU data, then plans collision-avoiding, kinematically feasible whole-body trajectories using its MRC-SDF model.The arm supplies joint-level control and real-time state feedback.Source: §III · p. 2

IV. HARDWARE DESIGN

  • QuadHand weighs approximately 1.7 kg and pairs a 330 mm-wheelbase quadrotor with a 3-DoF arm and gripper; its battery passively follows first-joint motion to keep CoG near the wheelbase center.This adjustment reduces disturbances to the aerial base.Source: §IV · p. 2

V. MULTI-ARTICULATED ROBOT-CENTRIC SDF

MRC-SDF represents the robot with preconstructed link-wise signed distance fields while retaining the environment’s original point-cloud geometry. This supports distance queries for whole-body configurations in close-range planning.

V. MULTI-ARTICULATED ROBOT-CENTRIC SDF

  • MRC-SDF associates each collision-checking link with a precomputed SDF, enabling distance queries between the UAM configuration and obstacles while preserving the original environmental point cloud.The representation is motivated by the trade-off between geometric accuracy and computational efficiency in cluttered, close-range interaction.Source: §V · p. 3
  • The representation preserves the environment’s original point-cloud geometry rather than replacing it with a geometric abstraction.Source: §V · p. 3
  • For each obstacle point, the method expresses its world-frame position in a link frame and obtains its signed distance to that link.Source: §V · p. 3

VI. TRAJECTORY OPTIMIZATION

The trajectory-optimization section formulates whole-body generation for dynamic feasibility and safety from free flight through task completion. It parameterizes the aerial base and manipulator jointly, with a flat output whose derivatives reconstruct system variables.

VI. TRAJECTORY OPTIMIZATION

  • The trajectory-generation problem considers dynamic feasibility and safety from free flight through task completion.Source: §VI · p. 3

A. Dynamic Model and Differential Flatness

  • The simplified UAM state includes base translation and rotation plus the manipulator’s joint angles and velocities, with three actuated joints excluding the gripper.The model includes arm-induced disturbance force and moment in base dynamics.Source: A. Dynamic Model and Differential… · p. 3
  • A flatness-based parameterization defines the flat output using base position, yaw, and joint angles, allowing system variables to be reconstructed from its derivatives.The manipulator joints are optimized jointly with the aerial-base trajectory variables.Source: A. Dynamic Model and Differential… · p. 3

B. Collision Avoidance

Collision avoidance evaluates obstacle points against link-wise SDFs while retaining the environment’s raw point cloud. The resulting differentiable safety cost supports whole-body trajectory optimization.

B. Collision Avoidance

  • The method queries link-wise SDFs for distances from each obstacle point to the UAM links, using the raw environment and target-object point cloud.Voxel-based or convex-hull representations are avoided because they lose geometric precision that may cause end-effector collisions.Source: B. Collision Avoidance · pp. 3–4
  • The safety cost penalizes configurations in which any part of the UAM approaches obstacles, using a predefined distance threshold and a smoothed ℓ1 penalty.The cost is evaluated over obstacle points and link-wise SDFs.Source: B. Collision Avoidance · p. 4
  • The safety-cost gradient is computed by the chain rule using link-SDF spatial gradients and forward-kinematics dependence on the UAM configuration.The formulation enables efficient, fully differentiable collision-cost evaluation while preserving raw-point-cloud geometric fidelity.Source: B. Collision Avoidance · p. 4

C. End-effector Insertion

An end-effector insertion penalty guides the two-finger gripper along a straight terminal segment toward its target pose. It combines position and orientation penalties and differentiates through forward kinematics.

C. End-effector Insertion

  • The insertion penalty is designed to facilitate stable straight-line insertion of the two-finger end-effector along the terminal segment to the target pose.The end-effector pose is computed from the whole-body pose and joint angles through the forward-kinematics chain.Source: C. End-effector Insertion · p. 4
  • The pose penalty sums smoothed-hinge position and orientation terms, with separate weights and tolerances for position and rotation.The rotation term uses the end-effector and goal rotation matrices.Source: C. End-effector Insertion · p. 4
  • The insertion penalty is differentiable with respect to the whole-body state through the forward-kinematics chain.The implementation uses gradients for position and rotation errors, including the position Jacobians with respect to the base and joint variables.Source: C. End-effector Insertion · p. 4

D. Dynamic Feasibility

Dynamic feasibility is enforced by penalizing violations in translational, rotational, and joint motion. Additional body-rate penalties keep x- and y-axis rotation within the feasible actuation range.

D. Dynamic Feasibility

  • A unified cost penalizes violations of translational velocity and acceleration, yaw rate, and actuated-joint angular velocity.These penalties enforce the aerial manipulator’s dynamic constraints for physical execution.Source: D. Dynamic Feasibility · p. 4
  • The method additionally penalizes body angular rates about the x- and y-axes to keep rotational motion within the feasible actuation range.The penalty uses the maximum allowable body angular velocity in the xy-plane.Source: D. Dynamic Feasibility · p. 5
  • The Hopf fibration formulation is used to simplify computation and avoid singularities when computing x- and y-axis body rates from the time derivative of the body z-axis.The body z-axis derivative can be computed from drone acceleration and jerk.Source: D. Dynamic Feasibility · p. 5

E. Trajectory Representation and Optimization Problem Formulation

The trajectory is parameterized with MINCO and optimized over waypoints and segment durations. Its objective combines trajectory smoothness, time regularization, and weighted costs for feasibility, safety, and insertion constraints.

E. Trajectory Representation and Optimization Problem Formulation

  • An s-order MINCO trajectory is a degree-(2s −1) polynomial that is continuously differentiable through order s−1 and uniquely optimizes the minimum-control problem for an s-integrator chain.The representation is compactly parameterized by intermediate waypoints and segment durations.Source: E. Trajectory Representation and… · p. 5
  • The optimization combines a polynomial-trajectory smoothness objective, total-time regularization, and weighted costs for dynamic feasibility, safety, and insertion constraints.The constraint costs include collision avoidance, end-effector insertion, and penalties on velocity, acceleration, yaw rate, joint rate, and xy-plane body angular velocity.Source: E. Trajectory Representation and… · p. 5
  • MINCO parameterizes the base position, yaw, and joint-angle trajectories; yaw and joint-angle initial guesses are linearly interpolated between initial and target configurations.The base-position trajectory is represented by its three coordinate components.Source: E. Trajectory Representation and… · p. 5

VII. EXPERIMENTS

Experiments assess figure-eight tracking and PCGC, then compare SFC, ESDF, and MRC-SDF for obstacle-course grasp planning.

A. Implementation Details

  • The onboard setup used a Jetson Orin NX and NxtPX4v2, with state estimated by an EKF fusing Vicon and IMU measurements.The simulation algorithm ran on an Intel Core Ultra 9 185H desktop CPU.Source: A. Implementation Details · p. 5

B. Figure-Eight Trajectory Tracking

  • With and without PCGC, tracking tests flew a 3.6 × 2 m figure-eight with arm-joint sinusoidal motion; 3D absolute errors were generally 3–4 cm, and PCGC improved accuracy.Each configuration included three flights; the payload case used two iron blocks totaling 150 g.Source: B. Figure-Eight Trajectory Tracking · p. 5

C. Simulation and Benchmark Comparisons

  • SFC corridors excluded the target region, ESDF at 0.02 m failed to resolve the cup gradient, while MRC-SDF retained finer geometry and the complete point cloud.The SFC exclusion made feasible grasping trajectories difficult; the ESDF resolution prevented the optimizer from finding a safe feasible trajectory.Source: C. Simulation and Benchmark Comparisons · pp. 5–6
  • In obstacle-grasping simulations, MRC-SDF achieved 99% success with 172.22 ms average optimization time under successful trials, the highest success rate among SFC and ESDF; ESDF achieved 0%.The task required passing two vertically placed window-shaped obstacles to grasp a cup on a cube from randomized initial states.Source: C. Simulation and Benchmark Comparisons · pp. 5–6

D. Real-world Implementation

Five real-world missions included general mobile manipulation and gap-crossing tasks, with initial states estimated online and trajectories computed onboard in real time.

D. Real-world Implementation

  • The five real-world missions included Drone Rescue, Cable Unplugging, and Plier Placement, alongside gap-crossing grasp and narrow-gap crossing.For all missions, the initial state was estimated online and the trajectory was computed onboard in real time.Source: D. Real-world Implementation · pp. 6–7, Figure 5 · p. 7

1) General Mobile Manipulation Tasks:

QuadHand performs mobile manipulation and gap-crossing tasks, and the conclusion describes its planning approach and current operating limits.

1) General Mobile Manipulation Tasks:

  • In three mobile tasks, QuadHand rescued a grounded Tello, unplugged a USB cable, and placed a pre-grasped plier on a hook.The three tasks were conducted without considering the environment point cloud.Source: 1) General Mobile Manipulation Tasks · p. 7

2) Gap-crossing Related Experiments:

VIII. CONCLUSION AND FUTURE WORK

  • The conclusion presents QuadHand’s compact quadrotor, 3-DoF arm, gripper, passive CoG compensation module, and MRC-SDF-based whole-body trajectory optimizer.The optimizer preserves raw point-cloud detail for close-proximity planning; experiments demonstrate safe aerial manipulation in complex environments.Source: §VIII · p. 7
  • Current limitations are reliance on external motion capture and manually predefined terminal states and retreat directions; future work targets onboard estimation and autonomous whole-body state generation and decision-making.Source: §VIII · p. 7
Loading 2609.35094v1…