Sequential Motion Planning for Bipedal Somersault via Flywheel SLIP and Momentum Transmission with Task Space Control

Xiaobin Xiong, Aaron D. AmesIROS 2020 · DOI

TL;DR

This paper presents a novel sequential motion planning and control framework for generating dynamic aerial maneuvers like somersaults on bipedal robots. By combining a simplified "Flywheel SLIP" model for ground phases with momentum transmission principles for flight, they enable robust planning and execution on a full-dimensional robot, demonstrating impressive somersaults in simulation.

What Problem Does This Paper Address?

Generating highly dynamic aerial motions, such as somersaults, on bipedal robots is a significant challenge. While impressive feats have been achieved by robots like Boston Dynamics' Atlas, these often rely on complex, full-dimensional models and trajectory optimization, which can be computationally expensive, prone to local minima, and offer little insight into the underlying mechanics. Furthermore, bipedal robots inherently have less stability margins compared to quadrupedal robots, making such dynamic maneuvers even more difficult. This paper aims to bridge this gap by developing a more principled and computationally tractable approach to plan and control somersaults, focusing on the essential dynamics that enable these maneuvers.

Methods

The core of the proposed method lies in a sequential planning approach that breaks down the somersault into distinct phases:

  • Ground Phase (Jumping & Landing): The authors introduce the Flywheel SLIP (FSLIP) model. This is a reduced-order representation that combines the dynamics of a Spring-Loaded Inverted Pendulum (SLIP) for the robot's hopping motion with an actuated flywheel attached to the mass. The flywheel captures the crucial angular dynamics needed for rotation. Trajectory optimization is performed on this simplified FSLIP model to generate desired jumping and landing trajectories, satisfying task constraints and physical limitations.
  • Flight Phase: During the flight phase, where the robot is airborne, the authors leverage the principle of conservation of centroidal angular momentum. Since the robot is underactuated in this phase, direct control is not possible. Instead, they propose momentum transmission from the flywheel to the lower body to achieve the desired rotational orientation. This allows for adjusting the robot's body orientation to prepare for landing.
  • Full-Dimensional Robot Control: The planned trajectories from the FSLIP model are then realized on a full-dimensional, 23-DOF bipedal robot (a modified Cassie). This is achieved using momentum-included Task Space Control (TSC). TSC formulates an optimization problem that drives the robot's actual outputs (e.g., momentum, leg length, joint angles) to match the desired outputs from the planning phase, while also considering physical constraints like ground reaction forces and motor torque limits.

Datasets

Not discussed in the paper. The work focuses on simulation-based validation using a modified Cassie robot model.

Key Insights and Ideas

  • Reduced-Order Modeling for Dynamic Maneuvers: The FSLIP model proves to be a powerful tool for capturing the essential dynamics of somersaults, enabling efficient trajectory optimization that would be intractable on a full-dimensional model.
  • Flywheel as a Key Component: The addition of a flywheel is crucial for generating the necessary angular momentum for somersaults. It acts as a proxy for the robot's upper body inertia and allows for controlled angular acceleration during the ground phase.
  • Momentum Transmission for Flight Control: Leveraging the conservation of angular momentum in the flight phase, the flywheel's momentum can be strategically manipulated to control the robot's body orientation, a critical step for preparing for landing.
  • Sequential Planning for Robustness: Breaking the somersault into distinct ground and flight phases, and then planning sequentially, allows for more robust control. The landing phase planning, for instance, occurs after the flight phase, making it less sensitive to uncertainties in the flight trajectory.
  • Practical Feasibility Challenges: The paper highlights that while the planning and control framework is sound, the practical realization on a physical robot hinges on powerful actuation and careful consideration of inertia distribution, particularly for the flywheel.

Key Results

The proposed method was implemented and validated in simulation on a modified Cassie robot. The authors successfully generated various somersault motions, including:

  • Frontflips with a 1m forward target distance.
  • Backflips.
  • Frontflips from a 0.5m platform.
  • A frontflip where the landing ground height was unknown (demonstrating robustness).

The simulation results show that the planned trajectories are tracked effectively by the full-dimensional robot, with successful landings and accurate forward distances, despite minor differences between the FSLIP model and the actual robot dynamics.

Limitations and Future Work

Limitations:

  • Actuation Requirements: The paper acknowledges that realizing these dynamic motions on a physical robot requires significant improvements in actuation design, particularly for the flywheel, which needs high speeds and potentially large torques.
  • Model Simplifications: While the FSLIP model is effective, it is a simplification. The paper notes that differences between the FSLIP and the full robot model can lead to discrepancies in jumping height.
  • Simulation-Based: The results are presented from simulation, and real-world deployment would introduce additional challenges.

Future Work:

  • Other Aerial Motions: Extending the framework to other dynamic aerial motions, such as periodic hopping, running, and aerial maneuvers in transversal and coronal planes.
  • Robots Without Free-Rotational Joints: Investigating how to achieve somersaults on robots that lack the same degree of freedom for rotation.
  • Control-Informed Robot Design: A key future direction is to unify mechanical design with control implementation, using insights from control requirements to inform robot design for dynamic aerial locomotion.

Summary updated

BibTeX
@inproceedings{xiong2020sequential,
  author = {Xiaobin Xiong and Aaron D. Ames},
  title = {{Sequential Motion Planning for Bipedal Somersault via Flywheel SLIP and Momentum Transmission with Task Space Control}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2020},
  doi = {10.1109/iros45743.2020.9341467},
}
Figure