Rolling Biped Polynomial Motion Planning

This work discloses a kinematic control model to describe the geometry of motion of a two-wheeled biped’s limbs. Limb structure is based on a four-bar linkage useful to alleviate damping motion during self-balance. The robot self-balancing kinematics geometry combines with user-customized polynomial vector fields. The vector fields generate safe reference trajectories. Further, the robot is forced to track the reference path by a model-based time-variant recursive controller. The proposed formulation showed effectiveness and reliable performance through numerical simulations.


Introduction
Motion planning is an essential function and a critical aspect in robotics engineering. Motion planning allows increasing the robot's degree of autonomy. Fundamentally a wide area of robotic tasks needs planning models such as: transportation and vehicular technology, service robotics, search, exploration, surveillance, biomedical robotic applications, spatial deployment, industry, and so forth. Moreover, motion planning is inherently impacted by the degree of holonomy and kinematic constraints in all robotic modalities: robot arms, legged robots, rolling platforms, marine/underwater vehicles, aerial robots, and including their end effectors. Depending on the robotic application, motion planning is designed either global or local. When the robot has an environmental map in advance, it is called global planning even with possibility to globally optimize routes. Alternatively, when there is only robot's local sensor data and the whole environment is unknown, it uses feedback from local observations. Planning methods can be generalized into four types: deterministic (based on mathematical numeric/analytic functions and models) [1][2][3], stochastic (recursive numerical methods based on probabilistic uncertainties) [4][5][6], heuristic (algorithms based on logical control and humanheuristic decision-making) [7][8][9][10], and mixed planning methods [11][12][13].
In this chapter, a kinematic motion/path planning method for path tracking of an inverted pendulum self-balancing rolling biped is deduced and discussed. This work is focused on the rolling biped motion modeling and simulation of the robot shown in Figure 1. The principal component of a rolling biped is self-balancing by controlling its pitch motion through in-wheel motors that allow rolling motion (inverted-pendulum-like). The robot's yaw motion is accomplished by the angular velocity resulting from the differential lateral speeds, which is a nonholonomic constrained model. The robot's design is purposed for missions to collect solid garbage in outdoors (a park), similar to other works [14]. Nevertheless, the robotic mission/task is out of this chapter's scope, instead a detailed geometry of motion is described in three parts: (i) motion planning for biped's balance, (ii) navigational path generation, and (iii) path tracking control. The work [15] proposed a balancing and dynamic speed control of a unicycle robot based on variable structure and linear quadratic regulator to follow a desired trajectory. The work [16] modeled a wheeled bipedal robot with analytic solutions of closed-form expressions in kinematic control loops. The work [17] reported a self-balancing two-wheeled robot with a manipulator on-board, using auto-balancing system to maintain force equilibrium. The work [18] applied a proportional integral derivative (PID) control and active disturbance rejection control to balance and steer a two-wheeled selfbalancing robot modeled by Lagrange formula. In [19], an adaptive robust control of a self-balancing two-wheeled underactuated robot to estimate uncertainty bound information, using deterministic system performance by Lyapunov method, was reported. In [20], a navigational two-wheeled self-balancing robot control using a PD-PI controller based on the Kalman filter algorithm was reported. Similarly, [21] used variable structure combining proportional integral differential controllers for balance and locomotion deriving a kinematic model based on the center of gravity constraint. Lagrangian-based with Kane's approach for dynamic balancing was reported in [22]. Moreover, [23] conducted a study using model predictive control for trajectory tracking of an inverted-pendulum wheeled robot. The work [24] reported a self-balancing robot controller using Euler-Lagrange and geometric control, and planar motion is controlled by logarithmic feedback and Lie group exponential coordinates. The work [25] developed a balancing and trajectory tracking system for an inverted-pendulum wheeled robot using a Lagrange-based backstepping structure variable method. This work's main contributions are an original design of limbs based on four-bar linkages to alleviate damping motion yielded from irregular terrains, from which a kinematic balancing condition is deduced. Further, polynomial vector fields with limit conditions are deduced from user-customized interpolation functions as path-generator models to yield safe routes in advance. Moreover, a recursive time-varying kinematic controller forces the robot to track resulting routes. The proposed system is demonstrated at the level of simulation. This chapter is organized in the following sections. Section 2 deduces the limb kinematics and its effects in the biped's balance. Section 3 presents the polynomial approach to trajectory generation by directional fields. Section 4 describes a navigation recursive controller for path tracking control. Finally, Section 5 presents the work's conclusions.

Balancing motion planning model
The main issue of an inverted-pendulum-like rolling biped is its capability to self-balance by controlling its pitch angle through the wheels velocity longitudinally (Figure 2(b)). This section deduces the balancing kinematics of the limb's planar linkage shown in Figure 2(a). As a difference from other biomechanical inspired muscle-tendon limbs [26], in this work each limb is comprised of a four-bar linkage operating as a double crank, where bars r and d are linked by a coupling link l with limited rotary angles.
The expressions provided in Proposition 2.1 are obtained from deductions in Appendix A. The passive joint point P 1 ¼ x 1 , y 1 À Á ⊤ mutually depends on the analytic model of the joint located at P 2 ¼ x 2 , y 2 À Á ⊤ to describe a rotary planar motion, proposed by Proposition 2.1 Limbs motion estimation. The four-bar linkage's passive joint P 2 is analytically described by.
Therefore, based on Definitions 4.1-4.3 of Appendix A, the model solution P 2 is From Eq. (56) deduced in Appendix A, there are two possible solutions for θ, described as opened or crossed motion. For this work, the type of motion should be ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi Q 2 À 4PR p ∀Q 2 ≤ 4PR that must be satisfied. Therefore, by using expressions (47), (52)-(56), the plots shown in 3 are obtained.
Furthermore, the wheels position is tracked by P f (Figure 2(a)). This design assumes the robot's center of gravity (cog) located at the same length f a a b (both in Figure 3) that is projected within the biped's body. Therefore, the cog is described by Definition 2.1: Definition 2.1 Robot's center of gravity (cog). The robot's cog is assumed to be located at Cartesian body's position, and Hence, in accordance to Definition 2.1, it follows that the robot's falling speed v f due to vertical unbalance is Taking into account that the robot is a dual differential drive kinematic structure, where v R and v L are the right-sided and left-sided velocities, respectively where s is the robot displacement over the ground, and r is the wheel's radius [m]. Likewise, the dual differential velocity is Hence, the robot's angular velocity ω rad=s ½ in terms of its differential speed v diff and constrained by its wheels lateral baseline distance b l and v diff is substituted to yield: The biped's falling angle and angular falling speed are λ rad ½ and _ λ rad=s which makes the robot's pitch turn around the wheel's center. Unbalancing The wheel's axis point moves at balancing speed v b that is parallel to the ground. Wheel's tangential speed is v w ¼ r _ φ, and linear motion speed is the balancing velocity underneath the biped's body at the wheel's axis, Therefore, the balancing condition is described by Definition 2.2: Definition 2.2 Kinematic balancing condition. The robot's vertical equilibrium condition is assumed for balancing and unbalancing velocities of equal magnitudes (or very approximated). Thus, and redefining, let _ φ B be the wheel's balancing angular speed, Therefore, it is of interest to find the wheel's rotary speed that balances the biped motion and from Definition 2.2, the following Proposition 2.2 arises, Proposition 2.2 Robot's balancing speed. The wheel's rolling velocity to satisfy the robot's body balance is proposed as and if and only if _ φ R ¼ _ φ L must exist, assuming that in such period of time, the robot's yaw ω t ≊0 allowing longitudinal balancing equilibrium. Thus, Thus, let us redefine the state variables as For an stability analysis for the balancing case and dropping off the highest-order derivative of the system The system total energy E k þ E p (kinetic plus potential) is a positive function, which is used as a Lyapunov candidate function, where the robot's translation motion r w _ φ B for equilibrium is and substituting _ φ B , For V x ð Þ ¼ E T and finding out that v(x) and _ V x ð Þ fulfill _ V 0 ð Þ ¼ 0. In addition, assuming that À2π ≪ λ ≪ 2π, then _ V 0 ð Þ > ∈ D À 0. Assuming that V : D !  is a continuous differentiable function. Finally, the following Lyapunov criterion is satisfied (20),

Polynomial vector fields
This section introduces the proposed path planning strategy to dynamically generate local paths. Two polynomial models are enhanced as vector fields to exert attractive and repulsive robot's accelerations. In principle, the desired acceleration functions are designed from the interpolating attractive/repulsive accelerations with respect to (w.r.t.) distance. The method used to fit points is the Lagrange formula: where the human-user establishes a desired numerical acceleration a i w.r.t. a desired distance δ i , from either goal or obstacle. Therefore, the following general cubic polynomial forms attractive a A and repulsive a R are obtained with their respective numeric coefficients λ i and β i , It follows that, Definition 3.1 establishes the acceleration path planning model toward a goal of interest. Definition 3.1 Planning toward a goal. Given the kinematic parameters λ 0 ¼ 0, Assuming that the distance between the goal and the robot δ g ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi x 2 þ y 2 2 p , such that x ≐ x g À x r and y ≐ y g À y r .
Similarly, Definition 3.2 establishes the acceleration path planning model that avoids obstacles zones. Definition 3.2 Planning obstacles avoidance. Given the kinematic parameters Assuming that the distance between the obstacle and the robot δ 0 ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi x 2 þ y 2 2 p , such that x ≐ x o À x r and y ≐ y o À y r . Figure 4(a) shows the robot's instantaneous acceleration toward a goal of interest. The attraction acceleration starts when the goal-robot distance δ g < 9m. The planner allows start from f A ≊0 to realistically provide speeds physically possible. Figure 4(b) shows the robot's instantaneous acceleration away from obstacles. The avoidance acceleration starts when the obstacle-robot distance δ o < 8m. This avoidance planner f R is faster than f A to increase confidence against obstacles.
Therefore, extending to two-dimension Cartesian space, let us deduce the following algebraic process for the goal-attraction planner f A : as well as Similarly, for the robot's acceleration to avoid obstacles, let us develop the following f R , by substituting the functional form of δ o into the gradient function subsequently by applying the gradient operator, ffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi ffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi Thus, by substituting terms in both local planners, the general attractive motion planning is a negative function with amplitude coefficient The attractive field is a positive function working within limits 0 ≤ δ g < 9, thus lim δ g !9 and the repulsive with The repulsive field is a negative function that works within range 0 ≤ δ o ≤ 8, The direction fields produced are shown in Figure 5, where for both cases the coordinates origin represents either goal or obstacle locations.
Moreover, when neither goals of interest nor obstacles are within the observation field of the robot, it must keep navigating along a prior route plan. The routing plan is map comprised of a sequence of Cartesian points g k ∈  2 , g k ¼ x, y ð Þ ⊤ .
When the robot accomplishes either f A or f R , it continues toward the following route point, expressed in terms of unit vectors: where a o is an ideal or averaged acceleration toward the next route point g 2 from g 1 . Therefore, the total path planner mode is given by Proposition 3.1.
Proposition 3.1 Total mission path planning. The total path planning model subjected to adaptive environmental changes is proposed as a vector fields sum: and the robot's instantaneous navigational total velocity 8

Motion Planning
Thus, to automatically limit the speed until the robot's maximal velocity v max , the real physical velocity v ph is constrained by the final planning motion (38) Figure 6 shows simulation results of the total navigation planner.

Path tracking
This section deduces an algebraic trajectory tracking linearized controller based on the kinematics geometry of the reference path: linear and angular velocities. Considering from previous Section 3, let v ph be called the reference velocity that is going to be tracked. Likewise, let (x 1,2 , y 1,2 ) be the next local planning coordinates to be reached (from location 1 to location 2). Hence, let us define x 1,2 ≐ x 2 À x 1 and y 1,2 ≐ y 2 À y 1 . Such that _ s 1,2 m=s ½ is a segment of the planning speed. Hence, let us obtain the first-order derivative, Similarly, by obtaining the desired robot's orientation θ 1,2 , its first-order derivative w.r.t. time is It follows that, the tracking control vector is Therefore, by stating (39) and (40) as a system of linear equations, the first-order derivatives must simultaneously be solved: thus, by factorizing both, the common term 1=s 1,2 and the first-order derivatives, the matrix form of the forward tracking law is Likewise, as the inverse tracking law is of our interest, it is inversely dropped off For notation simplicity, let us redefine _ ζ t ≐ _ x 1,2 , _ y 1,2 ⊤ and the time-variant control matrix K t as thus from Section 3, let u ref be the global reference planning model during time segment t 1,2 ≐ t 2 À t 1 such that, Therefore, the recursive path tracking control law is stated as whereû t is the instantaneous sensors observation of both components, displacementŝ and yawθ. The prediction control speed vector ζ tþ1 is the next desired local reference in line 2, whileζ t is the robot's Cartesian speed observer vector. Finally, u t + 1 is the control vector global prediction ( Table 1).

Conclusion
The mechanical design of the biped's lower-limb mechanical structure was configured as a double-crank four-bar linkage with passive-allowed motions. Motion planning began from determining the limbs' linkage positions causing the robot's height and pitch varying overtime producing unbalanced motions. Inferring balancing velocities to yield robot's vertical balance was possible and worked stable. The proposed balancing rolling condition was analyzed throughout its total energy model as a Lyapunov candidate function resulting stable in three criteria: V(0) = 0, v(x) = 0, and _ V x ð Þ ¼ 0. The proposed navigational approach allowed human-user to design short range-limited trajectories by cubic polynomials obtained from four empirical coordinates distance-acceleration by Lagrange interpolation. Both polynomial planning models have initial conditions a 0 t 0 ð Þ ¼ 0 m=s 2 ½ . Thus, the robot did not require an infinite acceleration to reach a desired speed at t 0 . Physically, the model is applied to any motion in equilibrium. For goal attraction the maximal acceleration is reached at the 55% of the distance, subsequently decreases monotonically until the goal position.
The navigational general planning model is a set of partial derivatives model combined, allowing dynamic local planning among multiple obstacles, goals, and routes. The navigational general planning model worked as the reference model for

Appendix A
This appendix provides the algebraic deduction of the limb's Cartesian motion model described in Section 2.