TL;DR
This paper presents a novel control architecture for bipedal robots that leverages their inherent compliance to achieve robust and efficient walking. By generating a library of compliant walking trajectories offline and using an inverse dynamics controller to track them online, the authors demonstrate successful walking on both simulated and real hardware, even on uneven terrain. This work is significant because it's the first to combine motion libraries with full-body dynamics and compliance for robots like Cassie.
What Problem Does This Paper Address?
Walking is a notoriously complex task for bipedal robots, especially those with compliant legs. While many approaches simplify the problem by modeling robots as rigid bodies (e.g., using Linear Inverted Pendulum models), this often sacrifices agility and energy efficiency. Incorporating compliance, which is crucial for natural and robust locomotion, introduces significant challenges: it increases the robot's underactuation, makes its dynamics more complex and sensitive to model inaccuracies, and complicates control design. Existing methods often struggle to achieve stable, efficient, and agile walking on compliant robots like Cassie, which has a complex, multi-link compliant leg structure. The authors aim to bridge this gap by developing a control strategy that explicitly accounts for and leverages this compliance.
Methods
The core of the paper's contribution lies in two main areas:
- Motion Library Generation: The authors create a library of compliant walking trajectories using the Hybrid Zero Dynamics (HZD) framework. This framework is well-suited for underactuated systems and allows for the generation of dynamic behaviors. Crucially, this library considers the full-body dynamics of the Cassie robot, including its compliance, which is a novel aspect compared to previous work that often ignored or simplified compliance. The trajectories are parameterized as Bézier polynomials.
- Inverse Dynamics Control: To track these generated trajectories in real-time, an orthogonal decomposition-based inverse dynamics controller is employed. This method is particularly effective for underactuated and floating-base robots. It decomposes the system's dynamics into constrained and unconstrained components, allowing for the computation of dynamically consistent feedforward torques. This controller is augmented with a standard PD feedback controller to correct for tracking errors. The authors also introduce velocity tracking and regulation heuristics to help the robot converge to a target walking speed and maintain stability.
Datasets
Not discussed in the paper. The authors generated their own trajectories via optimization and evaluated their controller on a physical robot and in simulation.
Key Insights and Ideas
- Compliance as an Advantage, Not Just a Challenge: Instead of trying to eliminate the effects of compliance, the authors show how to explicitly model and leverage it within the control framework. This allows for more natural and efficient movements.
- Offline Trajectory Optimization for Complex Dynamics: Generating a library of walking behaviors offline, parameterized by Bézier polynomials, is a practical approach to handle the high dimensionality and complexity of the problem. This allows for computationally intensive optimization to be done once, enabling faster real-time control.
- Orthogonal Decomposition for Robust Inverse Dynamics: This technique is key to handling the underactuation and floating-base nature of robots like Cassie, providing a stable way to compute feedforward torques.
- HZD Framework for Compliant Locomotion: The successful application of HZD to generate compliant walking trajectories demonstrates its power in creating dynamic and stable behaviors for robots with complex dynamics.
Key Results
- The proposed controller successfully enabled stable and robust walking on the Cassie robot in both simulation and on hardware.
- The system demonstrated the ability to track desired velocities and outputs, as shown in experimental results (Fig. 6).
- The controller was tested on flat terrain and over unplanned disturbances outdoors, indicating its robustness.
- The authors provide their controller and optimization source code on GitHub, facilitating further research.
Limitations and Future Work
- Trivial Double-Support Domain: The current model assumes a trivial double-support phase, which is difficult to achieve in practice with compliant legs. The authors suggest reintroducing a non-trivial double-support domain for enhanced stability.
- Analytical Inverse Dynamics Limitations: The current analytical inverse dynamics controller doesn't account for optimal distribution of contact forces, friction, or torque limits, nor does it provide explicit convergence guarantees.
- Future Control Development: The authors plan to focus on developing real-time model-based controllers, potentially by extending control Lyapunov function-based methods to walking. They also aim to incorporate more sophisticated handling of contact forces and constraints.
Summary updated