Nonlinear Model Predictive Control of a 3D Hopping Robot: Leveraging Lie Group Integrators for Dynamically Stable Behaviors

Noel Csomay-Shanklin, Victor D. Dorobantu, Aaron D. AmesICRA 2023 · DOI

TL;DR

This paper presents a novel approach to controlling a 3D hopping robot using nonlinear Model Predictive Control (MPC) integrated with Lie group theory. By geometrically aligning the control framework with the robot's rotational dynamics, they achieve stable 3D hopping and trajectory tracking on hardware, a first for online motion planning in this domain.

What Problem Does This Paper Address?

Controlling legged robots, particularly for dynamic behaviors like hopping, is a long-standing challenge. Hopping robots face intermittent dynamics, periods of underactuation, and very short ground contact phases where precise force application is crucial for regulating global state. While planar hopping has seen significant progress, 3D hopping control remains more complex. Existing methods often rely on simplified models or offline planning, limiting their ability to adapt to real-time changes or achieve highly dynamic maneuvers. This paper aims to leverage recent advancements in computational power and control theory to enable robust, dynamic 3D hopping using online motion planning.

Methods

The core of this work lies in a hybrid nonlinear MPC framework that respects the geometric structure of the robot's state space, particularly its orientation.

  • Hybrid System Modeling: The robot's dynamics are modeled as a hybrid system, switching between "flight" (in the air) and "ground" (contacting the floor) phases. This involves defining continuous dynamics for each phase and discrete "reset maps" for transitions, particularly during impacts.
  • Lie Group Integrators: To handle the rotational dynamics (represented by quaternions), the authors employ Lie group integrators. This ensures that control actions and state predictions remain geometrically consistent on the manifold of rotations, avoiding issues with singularities that plague other representations like Euler angles. They develop tools to linearize dynamics around a reference trajectory using Lie algebra elements.
  • Geometric MPC: The MPC problem is formulated to operate on these linearized dynamics, ensuring that the local approximations used by MPC are valid within the geometric structure of the state space. This involves constructing a sequence of linearized dynamics for each phase of the hybrid system.
  • Multi-Rate Architecture: A hierarchical control architecture is employed. A high-level MPC controller generates reference trajectories at 100 Hz, while a low-level quaternionic PD controller executes the control commands at 1 kHz, providing robustness to model errors and computational delays.

Datasets

Not discussed in the paper. The experiments are performed on a custom hardware platform (ARCHER) and in simulation (Mujoco).

Key Insights and Ideas

  • Geometric Consistency is Crucial for 3D Hopping: Treating rotational dynamics on their native Lie group manifold, rather than projecting them into Euclidean space, is essential for achieving stable and predictable 3D hopping. This geometric alignment prevents singularities and ensures that control actions respect the inherent structure of rotations.
  • Lie Group Integrators Enable Linearization: By using Lie group integrators and associated logarithmic maps, the authors can effectively linearize the robot's rotational dynamics around a reference trajectory. This is a key step in making nonlinear MPC tractable for complex, manifold-valued states.
  • Hybrid MPC for Dynamic Maneuvers: The combination of hybrid system modeling and MPC allows the robot to plan through discrete events like impacts, which are critical for hopping. The MPC can anticipate these events and plan control actions accordingly.
  • Low-Level Controller Complements MPC: The fast, low-level feedback controller provides robustness and handles rapid, high-frequency dynamics that might be challenging for the slower MPC loop to fully capture, especially during impact.

Key Results

  • Stable 3D Hopping on Hardware: The paper reports the first demonstration of sustained 3D hopping on a hardware platform using online motion planning.
  • Trajectory Tracking and Flipping in Simulation: The system successfully tracked global reference setpoints and performed dynamic maneuvers like flipping in simulation.
  • Robustness to Disturbances: The geometrically consistent approach demonstrated robustness to external disturbances.
  • Constant Velocity in Flight: Due to the underactuated nature of the robot in flight, the MPC implicitly plans for constant velocity in the air, relying on precise ground interactions for state regulation.

Limitations and Future Work

  • Impact Time Calculation: The paper notes that the impact time was calculated assuming an exactly vertical hopper, which is a simplification.
  • Foot Torque Simplification: The foot torque was set to zero in the optimizer, with the low-level controller handling foot compression.
  • Future Work: The authors suggest providing theoretical justifications for why predictive control can regulate underactuated systems, proving state and input constraint satisfaction, and incorporating more advanced features like planning with constrained footholds, friction cone constraints, and a higher-level decision-making layer. They also mention exploring ways to generate dynamically admissible trajectories within the inter-MPC time steps.

Summary updated

BibTeX
@inproceedings{csomayshanklin2023nonlinear,
  author = {Noel Csomay-Shanklin and Victor D. Dorobantu and Aaron D. Ames},
  title = {{Nonlinear Model Predictive Control of a 3D Hopping Robot: Leveraging Lie Group Integrators for Dynamically Stable Behaviors}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  doi = {10.1109/icra48891.2023.10160873},
}
Figure