Centroidal-momentum-based trajectory generation for legged locomotion
Introduction
Agile quadrupeds such as cats and squirrels are capable of performing highly dynamic maneuvers over a variety of terrains. When navigating in challenging environments, they can plan trajectories that fully utilize their physical capabilities and inherent dynamics. Such abilities, if successfully implemented on robots, possess profound potential in scenarios such as hazardous environment reconnaissance, disaster response and rescue.
Recent advances in quadruped robots has shown promising results in matching the dynamic capabilities of their natural counterparts: robust walking gaits have been performed by the BigDog [1], Spot and SpotMini from Boston Dynamics, though details of their control algorithms are yet to be published. ANYmal [2] is able to perform stable locomotion such as walking and trotting. Park et al. [3] displayed robust bounding motion on the MIT Cheetah 2 for a wide range of speeds in untethered 3D tests. Similar maneuverability has also been achieved on HyQ2Max [4] in creating trotting and self-righting while using hydraulic actuation. As the performance of the quadrupedal robots keeps advancing, so is the need for accurate algorithms capable of planning trajectories that fully utilize a robot’s dynamical capabilities to create agile motions, or to navigate challenging environments while respecting the robot’s physical restrictions.
The impediment to fast trajectory planning is partly rooted in the high degrees of freedom (DoF) present in robots. Methods based on a robot’s full-body dynamics have proven effectiveness in producing expressive motions, as shown in [5], [6], [7], [8], but would often suffer from long computational time. As a work-around, simplified models including the point-mass model [9] and other ones stemming from the linear inverted pendulum (LIP) model introduced by [10] have been used, such as the Reaction Mass Pendulum (RMP) [11] which augment the LIP models with a reaction-mass ellipsoid to capture the change in centroidal momenta for tasks such as balancing. A modified Spring Loaded Inverted Pendulum (SLIP) model is used by [12] to achieve running long jumps. Though less restrictive, these models still require task-specific modifications to its dynamic models to complement the full-body motions, or part of the dynamics will be left undecided or uncontrolled.
In recent years, the centroidal momentum (CM) of a robot, which is the aggregation of all links’ momenta at the robot’s Center of Mass (CoM) [13], has gained increasing attention as a reduced-order dynamic model in planning trajectories for robots with complex dynamics [14], [15], [16], [17]. A robot’s CM enjoys simple dynamics driven by the ground reaction forces (GRF) in legged locomotion. By including the GRF as optimization variables, a robot’s CM can be obtained through numerical integration, from which the corresponding joint angles can be recovered from kinematic equations without solving the complete second-order equation of motions. However, the use of CM dynamics in trajectory optimization for robots are often limited to a subsection of the optimization: either to validate transition feasibility between consecutive contact states [15], or to generate contact wrenches, which tends to rely on pre-defined contact locations and also requires additional process for joint trajectories [14]. Two notable exceptions come from the work of Dai et al. [17] and Fernbach et al. [15], where the contact states and forces are solved simultaneously as a hard linear complementarity problem. This method is shown to produce expressive motions and allows automatic gait scheduling, but often at the cost of long computation time when incorporating the full joint-space rigid body dynamics, or loss of generality
In this paper, we propose to augment the CM-dynamic-based trajectory optimization scheme by parameterizing the GRF and swing leg trajectories with Bézier polynomials. The optimizer simultaneously selects the Bézier polynomial coefficients and the contact leg’s joint trajectories, under an equality constraint which unifies the CM directly integrated from GRF through equations of motion, to the CM calculated from the robot’s generalized states and velocities using Centroidal Momentum Matrix [13], as shown in Fig. 1. This ensures that the obtained CM always respect the robot’s full-body dynamics, even between optimization nodes, unlike other collocation-based methods which approximate the solution of the model using polynomials. A similar approach can be found in [14], but is restricted to using a basic power basis polynomial for the CoM and GRF only, and relies on additional processing for the joint trajectories which requires prior knowledge of the exact contact locations.
The remainder of this paper is organized as follows: Section 2.1 describes the planar robot model used in formulating the proposed trajectory optimization algorithm. Section 2.2 presents two ways of obtaining a robot’s centroidal momentum with ground contact. Detailed formulation of the proposed multi-phase optimization framework, including transition map between different motion phases is presented in Section 3. Experimental validations are presented in Section 4 with reference trajectories obtained from the proposed method.
Section snippets
Simplified planar model
In this paper, a planar model shown in Fig. 2 is adopted to study a quadruped’s behavior in its sagittal plane, where the most of dynamic motions happen. The model consists of five massive links, with the torso link in the middle connecting two identical two-link legs through ideal revolute joints. The physical parameters of the planar model specified in Table 1.
The configuration of this five-link body can be described with its body angles:where subscripts F and H
Trajectory optimization
The values of a robot’s centroidal momentum derived in the section above should agree for any dynamically feasible trajectories. An equality constraint that unifies the two calculated values forms the basis of the proposed motion planning framework. This section presents the setup of this multi-phase trajectory optimization framework, followed by a test example to compare it with two methods
Experimental validation
This section introduces the planar testbed used in validating the proposed trajectory optimization framework, together with the experiment result of a 3-phase forward jumping motion. The optimization was based on the planar floating-base model described in Section 2.1 and formulated with the same structure proposed in Section 3.
Conclusion and future work
In this paper we present a trajectory optimization framework for planning legged locomotion based on a robot’s centroidal momentum, which enjoys simple dynamics dominated by the GRFs and the gravity. By parameterizing the GRFs as Bézier polynomials, the centroidal momentum of the robot can be solved analytically instead of through numerical integration. This avoids interpolating or curve-fitting the output trajectories between collocation points, which may result in sub-optimal trajectories or
CRediT authorship contribution statement
Chuanzheng Li: Conceptualization, Methodology, Software, Validation, Formal analysis, Writing - original draft. Yanran Ding: Investigation, Resources, Validation, Software. Hae-Won Park: Conceptualization, Methodology, Supervision, Project administration, Funding acquisition, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they do not have any financial or nonfinancial conflict of interests
Acknowledgment
This project is supported by NAVER LABS Corp. under grant 087387, Air Force Office of Scientific Research under grant FA2386-17-1-4665, and National Science Foundation under grant 1752262.
Chuanzheng Li received his B.S. degree in Mechatronics from Zhejiang University, Hangzhou, China in 2014, and the M.S. degree from the Mechanical Science and Engineering Department, University of Illinois at Urbana-Champaign, Champaign, IL, USA in 2017. He is currently in the Ph.D. program at University of Illinois at Urbana-Champaign supervised by Dr. Hae-Won Park, working primarily on the design of mechatronic systems and the real-time control of legged robots.
References (28)
- et al.
BigDog, the rough-terrain quadruped robot
IFAC Proc Vol
(2008) - et al.
SNOPT: An SQP algorithm for large-scale constrained optimization
SIAM Rev
(2005) - et al.
Anymal - a highly mobile and dynamic quadrupedal robot
2016 IEEE/RSJ International conference on intelligent robots and systems
(2016) - et al.
High-speed bounding with the mit cheetah 2: control design and experiments
Int J Rob Res
(2017) - et al.
Design of the hydraulically actuated, torque-controlled quadruped robot HyQ2Max
IEEE/ASME Trans Mechatron
(2017) Using optimization to create self-stable human-like running
Robotica
(2009)- et al.
Synthesis and stabilization of complex behaviors through online trajectory optimization
2012 IEEE/RSJ International conference on intelligent robots and systems
(2012) - et al.
A direct method for trajectory optimization of rigid bodies through contact
Int J Rob Res
(2014) - et al.
Trajectory optimization through contacts and automatic gait discovery for quadrupeds
IEEE Rob Autom Lett
(2017) - et al.
Single leg dynamic motion planning with mixed-integer convex optimization
2018 IEEE/RSJ International conference on intelligent robots and systems
(2018)
The 3d linear inverted pendulum mode: a simple modeling for a biped walking pattern generation
2001 IEEE/RSJ International conference on intelligent robots and systems
Reaction mass pendulum (RMP): an explicit model for centroidal angular momentum of humanoid robots
2007 IEEE/RSJ International conference on robotics and automation
Development of high-span running long jumps for humanoids
2014 IEEE International conference on robotics and automation
Centroidal momentum matrix of a humanoid robot: Structure and properties
2008 IEEE/RSJ International conference on intelligent robots and systems
Cited by (14)
Modeling and simulation for designing a line walking chameleon-like legged robot
2022, Simulation Modelling Practice and TheoryCitation Excerpt :In kinematics-based approaches, the key idea is to analyze the forward and inverse kinematics of the legs and to coordinate a sequence of steps (typically using state machines) for producing a gait [20]. Examples of this approach can be found in [19,21–25]. In the bionspired-based approach, also known as biomimetic control architectures, the control strategy is inspired by the central pattern generators (CPGs), which are primarily responsible for generating coordinated and rhythmic locomotion patterns of animals [20,26].
Kinematics of the center of mass for robotic mechanisms based on lie group theory
2022, Mechanism and Machine TheoryCitation Excerpt :David E. Orin et al. presented centroidal dynamics for bipedal robot with all limbs mass [9], significantly reduced unnecessary trunk bending during balance maintenance against external disturbance. Chuanzheng Li et al. used this method to complement the quadruped robot motions [10]. Mobile manipulating robot is widely used in industry and exploration fields, which has a mobile chassis with wheels [11–13] or legs [14–16] and one or two manipulators fixed on its chassis.
Learning Footstep Planning for the Quadrupedal Locomotion with Model Predictive Control
2022, Lecture Notes in Networks and SystemsRepresentation-Free Model Predictive Control for Dynamic Motions in Quadrupeds
2021, IEEE Transactions on RoboticsHybrid Parallel Compliance Allows Robots to Operate With Sensorimotor Delays and Low Control Frequencies
2021, Frontiers in Robotics and AI
Chuanzheng Li received his B.S. degree in Mechatronics from Zhejiang University, Hangzhou, China in 2014, and the M.S. degree from the Mechanical Science and Engineering Department, University of Illinois at Urbana-Champaign, Champaign, IL, USA in 2017. He is currently in the Ph.D. program at University of Illinois at Urbana-Champaign supervised by Dr. Hae-Won Park, working primarily on the design of mechatronic systems and the real-time control of legged robots.
Yanran Ding received B.S. degree from UM-SJTU joint institute, Shanghai Jiao Tong University, Shanghai, China in 2015. He has received the M.Sc. degree from the Mechanical Science and Engineering Department, University of Illinois at Urbana-Champaign, Champaign, IL, USA in 2017, and he is currently working towards his Ph.D. degree at the Dynamic Robotics Laboratory. His research interests include the design and control agile robotics systems. His research include mechanism design, actuation systems and concentrates on using applied optimization and control theory to enable legged robots to achieve dynamic motions. He is one of the best student paper finalists in IROS 2017.
Hae-Won Park is an Assistant Professor of Mechanical Engineering at the Korean Advanced Institute of Science and Technology (KAIST). He received B.S. and M.S. degrees from Yonsei University, Seoul, Korea, in 2005 and 2007, respectively, and the Ph.D. degree from the University of Michigan, in 2012, all in mechanical engineering. His research interests lie at the intersection of control, dynamics, and mechanical design of robotic systems, with special emphasis on legged locomotion robots. He is the recipient of the 2018 National Science Foundation (NSF) CAREER Award, NSF’s most prestigious awards in support of early-career faculty.