Pedipulation in Quadruped Robots: Toward Autonomous Model-Based Legged Manipulation

Pedipulation in Quadruped Robots: Toward Autonomous Model-Based Legged Manipulation Marco Tabita1, Carmine T. Recchiuto1, Enrico Simetti1, and Antonio Sgorbissa1 Abstract—Quadruped robots are increasingly used for exploration, inspection, and monitoring in environments where wheeled or tracked vehicles are inadequate. While locomotion has been extensively studied, the use of legs not only to achieve locomotion but also to manipulate the environment is largely unexplored. Inspired by biological counterparts, pedipulation can extend autonomy by enabling non-prehensile interactions such as pushing obstacles or opening doors, as well as simple social gestures that may enhance acceptance in human-centered environments. This paper presents a modular task-priority control framework for pedipulation on the Unitree Go1 quadruped. The architecture combines inverse dynamics, balance maintenance, and mission-level sequencing, allowing the robot to execute pedipulation tasks while preserving stability. The framework was validated in simulation and real-world experiments across representative scenarios, including object pushing, door opening, object removal, and human interaction. Index Terms—quadruped robots, pedipulation, task-priority control, model-based control I. INTRODUCTION Quadruped robots stand out among mobile platforms for their balance of agility, stability, and compact dimensions, which makes them suitable for specialized tasks such as exploration, search and rescue, and monitoring across different environments. To fully realize their potential and more closely emulate their biological counterparts, these robots require advanced locomotion capabilities combined with manipulation skills. Early efforts to extend manipulation involved mounting an additional arm on the quadruped’s back, but this approach increases weight, complexity, and energy consumption of the overall system. A more efficient alternative is to exploit the legs themselves for manipulation, a concept known as pedipulation [6], [3]. This allows quadrupeds to interact with their surroundings not only for locomotion but also for tasks such as pushing obstacles, pressing buttons, or opening doors, while broadening their role in exploration and human-centered environments. Pedipulation, however, introduces unique challenges. The robot must maintain balance while using one leg for manipulation, which requires control strategies capable of managing dynamic interactions and ensuring smooth transitions between locomotion and manipulation. Early loco-manipulation approaches, typically with an additional arm, addressed control through hierarchical task formulations [4], [7]. These frameworks ensure stability and precision by structuring objectives 1All authors are with the DIBRIS Department, University of Genoa, Via All’Opera Pia 13, 16145 Genoa, Italy. Corresponding author: [email protected] in order of priority, but compute solutions online without longterm optimization, a limitation that Model Predictive Control (MPC) overcomes by optimizing trajectories over a temporal horizon [2], [4]. Pedipulation itself has been mostly explored through model-free methods [6], [9], with some exceptions where model-based low-level controllers are combined with learning-based planners [3]. However, these approaches do not exploit the full-body dynamics of the quadruped; they rely on impedance control for the pedipulating leg. Our goal instead is to formulate a model-based whole-body low-level controller to exploit all degrees of freedom of the classic quadrupedal robot, by leveraging the properties of a whole-body task priority controller coupled with MPC controller. In this work, we present a preliminary modular task-priority control framework for executing pedipulation on the Unitree Go1. The architecture integrates inverse kinematics to compute feasible foot trajectories, together with dynamic tasks, balance maintenance, and mission-level sequencing, enabling the robot to perform pedipulation while preserving stability. The framework was validated in both simulation and real-world experiments across representative scenarios. II. METHODOLOGY The proposed pedipulation framework is structured around a modular control architecture (Figure 1). Its key components include the Main Loop, the Mission Manager, and an inverse kinematics (IK) solver, together with the Unitree IO-Interface, the Unitree State Estimator, and a robot model that provides dynamic and kinematic data. At its core, the system employs a task-priority controller, where control objectives are formulated as equality or inequality tasks organized hierarchically and solved as a Hierarchical Quadratic Program (HQP). For pedipulation, the most relevant tasks are: (i) unactuated base dynamics constraint, (ii) balancing constraints ensuring stability, (iii) ground contact handling, (iv) body posture control, and (v) leg posture for the pedipulating leg. Implemented using the TSID library [10] these the QP solution gives the optimal joint accelerations and contact forces to achieve the desired tasks while respecting physical constraints. The final torque commands are computed using inverse dynamics. Trajectory references for the body and pedipulating foot are generated using fifth-order polynomials with zero initial and final velocity and acceleration. At this stage, to test the task-priority framework itself, trajectories for the leg are computed in joint space using a simple IK solver. This choice avoids inconsistencies with the kinematic constraints of the leg, which made Cartesian trajectories difficult 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629766 137 Unitree State Estimator Unitree IO-Interface HQP Controller Main Loop IK Solver Mission Manager Robot Model /joint_states /odom Phase 1 Phase N Mission Definition Fig. 1. Control framework architecture: the structure that collects all modules is the Main Loop. The Unitree IO-Interface and the Unitree State Estimator are provided by the Unitree repository. to track reliably. In future work, the framework is intended to be coupled with a MPC controller to generate feasible wholebody trajectories that exploit all available degrees of freedom. The Mission Manager orchestrates pedipulation behaviors by sequencing phases such as body motion, attitude adjustment, leg transition between support and manipulation, and target reaching. By combining these primitives, a wide range of pedipulation tasks can be executed while maintaining stability and robustness. To evaluate the proposed framework, a set of missions was defined, each involving different pedipulation tasks inspired by scenarios commonly encountered in daily routines. These tasks included: (i) pushing objects, where obstacles can sometimes be managed simply by moving them aside, while in other cases the robot may encounter a closed door that must be opened to proceed; (ii) removing objects from the robot’s back, since during ordinary routines objects may accidentally fall onto the robot and pedipulation can be used to remove them, provided joint limits allow the required motions; (iii) obstacle sensing, where the robot’s paws, if equipped with suitable sensors such as the pressure sensors available on the Unitree Go1, can be used not only to detect contacts but also to analyze potential obstacles encountered during exploration, assuming sufficient sensitivity and proper calibration; and (iv) interacting with humans, which is crucial for social acceptance and effective collaboration in human-centered environments. All defined missions were successfully completed, with the robot effectively using its four legs to perform pedipulation tasks as required. Specific tasks performed in real scenarios included opening an office door, with a foot mean positioning error of 0.059 m; interacting with a cardboard cat placed around the robot (in front, behind, and on its back), with mean positioning errors of 0.032 m, 0.031 m, and 0.075 m, respectively; interacting with a human, with a mean error of 0.062 m; and reaching various targets to inspect soft and fixed obstacles, with a mean error of 0.041 m. III. CONCLUSION We introduced a preliminary model-based framework for pedipulation in quadruped robots based on a modular taskpriority controller. The architecture combines inverse kinematics, dynamic tasks, balance maintenance, and mission-level sequencing, enabling the Unitree Go1 to perform pedipulation while preserving stability. Validation in simulation and realworld experiments, including obstacle pushing, door opening, object removal, and social gestures, confirmed feasibility, with positioning errors confined to a few centimeters. The current framework remains sensitive to initial conditions and lacks long-horizon trajectory optimization, since trajectories are generated using fifth-order polynomials in Cartesian space for the body and a joint-space for the pedipulating leg. Future work will address these limitations by coupling the controller with Model Predictive Control for dynamically consistent whole-body trajectories, and by enhancing state estimation and perception to support autonomous and adaptive behaviors in real-world environments. REFERENCES [1] Sombolestan, M., et al. ”Hierarchical Adaptive Loco-manipulation Control for Quadruped Robots” 2022. [2] Sleiman, A., et al. ”A Unified MPC Framework for Whole-Body Dynamic Locomotion and Manipulation” 2021. [3] He, Y., et al. ”Learning Visual Quadrupedal Loco-Manipulation from Demonstrations” 2024. [4] Grandia, R., et al. ”Perceptive Locomotion through Nonlinear Model Predictive Control” 2022. [5] Farshidian, F., et al. ”Real-Time Motion Planning of Legged Robots: A Model Predictive Control Approach” 2018. [6] Arm, C., et al. ”Pedipulate: Enabling Manipulation Skills using a Quadruped Robot’s Leg” 2024. [7] Bellicoso, C. D., et al. ”ALMA-Articulated Locomotion and Manipulation for a Torque-Controllable Robot” 2019. [8] Bellicoso, C. D., et al. ”Perception-less Terrain Adaptation through Whole Body Control and Hierarchical Optimization” 2016. [9] Cheng, L., et al. ”Legs as Manipulator: Pushing Quadrupedal Agility Beyond Locomotion” 2023. [10] Del Prete, A., et al. ”Implementing Torque Control with High-Ratio Gear Boxes and without Joint-Torque Sensors” 2016. 138