TL;DR
This paper presents a novel method for controlling bipedal robots to hop by embedding a simplified "spring-mass" model into the robot's full dynamics. This approach allows for computationally efficient planning of hopping trajectories and synthesis of a robust, optimization-based controller that has been successfully demonstrated on the Cassie robot.
What Problem Does This Paper Address?
Controlling dynamic behaviors like hopping on legged robots is challenging due to their inherent complexity and underactuation. Traditional approaches often rely on simplified models (like the spring-mass model) for planning but struggle to faithfully translate these plans to the full-order robot dynamics without significant computational overhead or loss of fidelity. This gap makes it difficult to achieve agile and robust dynamic motions. The paper aims to bridge this gap by creating a controller that effectively embeds the insights from a simplified model into the complex reality of a physical robot.
Methods
The core of the paper's methodology lies in two key areas:
- Reduced-Order Model Identification and Embedding:
- A spring-mass model is derived from the Cassie robot's physical properties, specifically its leg compliance. The leg's stiffness and damping are modeled as a prismatic spring, and this spring's properties are derived as functions of the robot's leg length.
- This spring-mass model is then embedded into the full-order robot dynamics. Instead of directly controlling the robot's joints, the controller aims to make the robot's actual leg length and other outputs track trajectories generated by the optimized spring-mass model.
- Optimization-Based Control Synthesis:
- Trajectory Optimization: A direct collocation method is used to plan optimal jumping and landing trajectories for the spring-mass model, minimizing energy consumption while achieving desired heights and stable landings.
- Control Lyapunov Function-based Quadratic Program (CLF-QP): The planned leg length trajectories (and other outputs like centroidal momentum) are used as desired outputs for a CLF-QP controller. This formulation synthesizes a nonlinear feedback control law that drives the robot's actual outputs to track the desired trajectories, ensuring stability. The CLF-QP formulation is extended to handle the hybrid nature of hopping (jumping, flight, landing phases) and incorporates constraints like ground reaction forces and torque limits.
Datasets
Not discussed in the paper. The paper focuses on deriving a model from the robot's physical properties and validating it through simulation and experiments on the Cassie robot.
Key Insights and Ideas
- Faithful Embedding of Reduced-Order Models: The key insight is that by deriving the spring-mass model's parameters (stiffness, damping) directly from the robot's compliant components and then using the leg length as a "virtual actuator" in the reduced-order model, the planned trajectories can be more faithfully embedded into the full-order robot.
- Leg Length as a Control Primitive: Treating leg length as a primary output to be tracked by the controller, rather than just a kinematic variable, allows for a more direct translation of the spring-mass model's dynamics into the robot's control.
- CLF-QP for Hybrid Systems: The application of CLF-QP to a multi-domain hybrid system (jumping, flight, landing) is a powerful technique for synthesizing controllers that can handle distinct phases of motion and transitions between them.
- Centroidal Angular Momentum for Stability: Including centroidal angular momentum as an output to be stabilized during the flight phase is crucial for preventing unwanted whole-body rotations in an underactuated system.
Key Results
- The proposed framework was successfully implemented and validated on the 3D bipedal robot Cassie in both simulation and real-world experiments.
- The robot achieved hopping with approximately 7 inches of ground clearance and an air-time of about 0.423 seconds.
- The ground reaction forces and toe-off timing in the experimental hopping closely matched those predicted by the spring-mass model, indicating a faithful representation and control strategy.
- The CLF-QP controller demonstrated fast convergence of the robot's outputs to the desired trajectories.
Limitations and Future Work
- Model Inaccuracy: The paper acknowledges that model inaccuracies of the physical robot led to the robot jumping forward slightly more than intended in the experiments. Future work will involve addressing this by utilizing feedback to further refine the control.
- Future Work Directions:
- Extending the approach to other dynamic behaviors like walking and running in 2D and 3D.
- Establishing theoretical stability conditions that quantitatively link the reduced-order model's predictions to the full-order system's dynamics.
- Developing more sophisticated controllers for complex robots by leveraging simple models.
Summary updated