The pursuit of creating machines that can traverse the world with the grace, efficiency, and adaptability of animals has long been a central drive in robotics. Among various configurations, quadrupedal platforms stand out for their inherent stability, superior load-bearing capacity, and potential for high-speed locomotion across unstructured terrain. My research is fundamentally rooted in this pursuit, focusing on imbuing a bionic robot with the organic movement principles found in nature. The core challenge, and the focus of this extensive discussion, lies not merely in the mechanical replication of a limb but in the intelligent planning of its path through space—the foot trajectory. This article details my first-person journey in developing a biomimetic foot-end trajectory planning strategy, moving beyond conventional function-based methods to one directly inspired by the kinematics of a living mammal.
The design philosophy for my bionic robot began with a close observation of canine physiology. The skeletal structure, joint arrangement, and muscle coordination of a large dog served as the blueprint. The leg mechanism was designed as a fully articulated, serial-link manipulator, mirroring the biological counterpart to ensure a similar range of motion and kinematic compatibility. However, possessing a biologically-inspired skeleton is insufficient; it requires a soul—a gait. Traditional trajectory planning methods, such as cycloidal, elliptic, or simple polynomial functions defined in Cartesian space, are computationally straightforward. Yet, they often result in motions that appear mechanical, can generate significant impact forces upon foot landing, or are not optimally efficient. My goal was to derive a trajectory that was intrinsically natural, smooth, and adaptable.
The foundational step was to acquire genuine biological motion data. I employed a vision-based motion capture approach using a depth sensor capable of skeletal tracking. The subject was a large dog trained to walk at a steady pace on a flat surface using a trotting gait (Trot), characterized by diagonal leg pairs moving in unison. The sensor was positioned perpendicular to the plane of motion, and the three-dimensional coordinates of key skeletal joints were recorded throughout multiple complete gait cycles.
From this skeletal data stream, I isolated the trajectory of a single foot—specifically, the right hind foot—over one complete stride. The pixel coordinates from the video stream were transformed into a real-world coordinate system with the origin at the initial contact point. This process yielded a discrete set of $n$ data points $P_i = (x_i, y_i)$, representing the foot’s path relative to the body during the swing and stance phases. A subset of this coordinate data is presented in the table below to illustrate the spatial progression.
| Point Index | x (cm) | y (cm) |
|---|---|---|
| 1 | 0.00 | 0.00 |
| 10 | 12.37 | 2.37 |
| 20 | 23.35 | 8.50 |
| 30 | 27.96 | 9.99 |
| 40 | 32.90 | 6.18 | 51 | 38.50 | 0.00 |
With this cloud of discrete points representing the biological footfall, the task was to generate a continuous, smooth trajectory suitable for driving the joints of my bionic robot. The criteria were stringent: the resulting curve must be smooth (C² continuous to ensure finite acceleration), must closely approximate the biological data, and must exhibit properties conducive to stable and efficient robotic locomotion, such as near-zero vertical velocity at touch-down and lift-off. I evaluated and compared two powerful curve-fitting techniques: Cubic B-splines and Cubic Bézier curves.
Cubic B-spline Fitting: B-splines are renowned for their local control and excellent smoothness properties. For a set of control points ${C_0, C_1, …, C_n}$, a cubic B-spline segment $B_i(u)$ between nodes is defined parametrically. To ensure the curve starts at $C_0$ and ends at $C_n$, I employed a standard technique of extending the control polygon with phantom points. The $i$-th segment of a uniform cubic B-spline is given by:
$$
B_i(u) = \frac{1}{6} \begin{bmatrix} u^3 & u^2 & u & 1 \end{bmatrix}
\begin{bmatrix}
-1 & 3 & -3 & 1\\
3 & -6 & 3 & 0\\
-3 & 0 & 3 & 0\\
1 & 4 & 1 & 0
\end{bmatrix}
\begin{bmatrix}
C_i\\
C_{i+1}\\
C_{i+2}\\
C_{i+3}
\end{bmatrix}, \quad u \in [0,1]
$$
This formulation guarantees continuity in position, velocity, and acceleration across segment boundaries, which is crucial for preventing jerky motion in the bionic robot.
Cubic Bézier Fitting: As a comparison, I also fitted the data using piecewise cubic Bézier curves. A Bézier curve is defined by four control points $P_0, P_1, P_2, P_3$:
$$
P(\tau) = (1-\tau)^3 P_0 + 3(1-\tau)^2\tau P_1 + 3(1-\tau)\tau^2 P_2 + \tau^3 P_3, \quad \tau \in [0,1]
$$
While Bézier curves interpolate their endpoints ($P_0$ and $P_3$), achieving global smoothness requires constraining the intermediate control points of adjacent segments, which offers less local flexibility than B-splines for fitting a complex path like a foot trajectory.
I implemented both fitting algorithms computationally. The primary metrics for evaluation were the Root Mean Square Error (RMSE) of the fit, the maximum deviation, and critically, the kinematic properties of the generated trajectory (velocity and acceleration profiles). The quantitative comparison of the fitting performance is summarized below:
| Fitting Method | RMSE (cm) | Max Error (cm) | Min Error (cm) | Global C² Continuity |
|---|---|---|---|---|
| Cubic B-spline | 1.04 | 3.36 | 0.01 | Yes |
| Piecewise Cubic Bézier | 1.12 | 9.32 | 0.00 | No (Requires constraints) |
The analysis was decisive. While the Bézier fit technically passed through more data points (min error of 0), the Cubic B-spline fit produced a lower overall RMSE and a drastically smaller maximum error. More importantly, the B-spline curve possessed inherent global smoothness without additional constraints, a vital property for the bionic robot’s motion controller. The kinematic derivatives of the B-spline trajectory confirmed its superiority. The velocity and acceleration profiles at the foot-point were smooth and, crucially, the vertical ($y$-direction) velocity and acceleration approached zero at the moments of ground contact and lift-off. This characteristic is paramount for a bionic robot as it minimizes impact forces, reduces slipping, and leads to more stable and energy-efficient locomotion.
$$
v_y(t_{touchdown}) \approx 0, \quad a_y(t_{touchdown}) \approx 0
$$
The next step was to translate this optimal Cartesian foot trajectory into actionable commands for the bionic robot’s leg actuators. Using the inverse kinematics model of the robot’s 2-DOF leg (comprising hip and knee joints), I mapped the smooth $(x(t), y(t))$ trajectory from the B-spline into the corresponding joint angle trajectories $\theta_{hip}(t)$ and $\theta_{knee}(t)$. The resulting joint profiles were smooth and remained well within the mechanical limits of the servo actuators, proving the physical feasibility of the biomimetic path for my specific bionic robot design.
A critical advantage of a bionic robot is its potential for energy efficiency, mirroring the optimized movement of animals. To evaluate this, I developed an energy consumption model for the electrically-actuated leg joints. The instantaneous electrical power for the $j$-th joint on the $i$-th leg is a function of its torque $\tau_{ij}(t)$ and angular velocity $\dot{\theta}_{ij}(t)$, accounting for motor and drivetrain efficiencies ($\eta$), resistance ($R$), and gear ratios ($\xi$). The total energy $E$ over a gait cycle $T$ can be approximated by integrating the power losses:
$$
E \approx \sum_{i=1}^{4} \sum_{j=1}^{2} \int_{0}^{T} \left( \frac{\Delta(P_{m,ij}(t))}{\eta} + \frac{R \tau_{ij}^2(t)}{\psi^2 \eta^2 \xi^2} \right) dt
$$
where $\Delta(x)=x$ if $x>0$ else $0$ (considering mostly motoring, not regenerative braking), $P_{m}$ is mechanical power, and $\psi$ is the motor torque constant. Simulation of this model for the biomimetic B-spline trajectory showed a favorable, progressive power profile without spikes, indicating a gait that is not only stable but also efficient for the bionic robot platform, promoting longer operational periods on a single battery charge.
The final validation involved full-body dynamics simulation. Implementing the biomimetic foot trajectory for all four legs in a coordinated trot gait, I analyzed the motion of the robot’s center of mass (CoM). The results were compelling. The CoM trajectory in the forward direction was nearly linear with constant velocity, indicating steady progression. The lateral and vertical oscillations of the CoM were minimal, demonstrating inherent gait stability. This is a direct consequence of the smooth, biologically-inspired foot placement which manages momentum effectively.
| Performance Metric | Biomimetic B-spline Trajectory | Traditional Cycloidal Trajectory |
|---|---|---|
| Foot Impact at Touchdown | Very Low (Near-zero vertical velocity) | Moderate (Non-zero vertical velocity) |
| Trajectory Smoothness | High (C² continuous) | High (C² continuous) |
| Approximation to Biological Data | Excellent (Low RMSE) | Poor (Not based on bio-data) |
| Center of Mass Stability | High (Low vertical/lateral oscillation) | Moderate |
| Energy Efficiency (Simulated) | High (Smooth power profile) | Lower (Due to higher impacts) |
In conclusion, this first-person research narrative demonstrates a comprehensive methodology for bestowing a bionic robot with a genuinely natural gait. By directly observing and digitizing mammalian locomotion, processing the data through an optimal curve-fitting framework (Cubic B-spline), and rigorously analyzing the kinematic, dynamic, and energetic outcomes, I have developed a trajectory planning strategy that transcends conventional methods. The resulting motion is not just a sequence of positions; it is a smooth, stable, and efficient transfer of principles from biology to engineering. This biomimetic approach provides a robust foundation for the bionic robot. Future work seamlessly extends from here: adapting this core method to plan trajectories for climbing stairs, traversing rocky terrain, or dynamically adjusting stride length and height by modulating the original biological dataset or the control points of the spline. The path forward for advanced bionic robot locomotion is clear—by looking more deeply at nature, we can teach our machines to move with a life-like proficiency that was previously the sole domain of animals.