Versatile Real-Time Motion Synthesis via Kino-Dynamic MPC With Hybrid-Systems DDP

Wenhao Yu, He Li, Tingnan Zhang, Wenhao Yu 0003, Patrick M. Wensing

TL;DR

This paper introduces a unified Model Predictive Control (MPC) framework that allows quadruped robots to seamlessly switch between regular gaits and complex, agile motions like jumping, all in real-time. By employing a novel hybrid kinodynamic model and a specialized Differential Dynamic Programming (DDP) solver, the system achieves unprecedented versatility and robustness on hardware platforms.

What Problem Does This Paper Address?

Quadruped robots are incredibly agile in the real world, capable of a wide range of movements from walking and trotting to dynamic jumps and rapid transitions between these skills. However, replicating this versatility in robots has been a significant challenge. Existing approaches often rely on separate controllers for regular gaits and specialized maneuvers like jumping. This leads to a fragmented system that struggles with seamless transitions and on-the-fly adaptation. The core problem is achieving a unified, real-time control system that can synthesize a diverse set of agile motions, including complex maneuvers and their transitions, without pre-computed trajectories.

Methods

The paper's core contribution is a Hybrid Kinodynamic Model Predictive Control (HKD-MPC). This approach combines:

  • Hybrid Kinodynamic Model: This model extends the standard Single Rigid Body (SRB) model by incorporating leg kinematics. Crucially, it dynamically switches how leg contact is modeled: in stance, the foot is fixed; in swing, joint velocities are controlled. This allows for a more accurate representation of robot dynamics during complex maneuvers.
  • Differential Dynamic Programming (DDP) Solver: The MPC problem is solved using a constrained variant of DDP, specifically a Hybrid-System DDP (HS-DDP). This solver is designed to handle the state-based switching inherent in the hybrid model.
  • Online Re-planning with Feedback: To achieve real-time performance, the authors employ two key schemes:
  • Feedback Gains: Instead of just using the feedforward control from the previous optimization, they also incorporate the feedback gain matrix from the DDP solver. This "feedback MPC" significantly improves robustness and convergence during re-planning, especially after disturbances like touchdowns.
  • Re-initialization of HS-DDP: For subsequent re-planning steps, they reuse the Lagrange multipliers from the previous solution while resetting the penalty coefficient. This strategy helps maintain fast convergence, requiring only a few DDP iterations per re-planning step.
  • Unified Control Architecture: The HKD-MPC is integrated into a broader framework that includes a gait library, reference generation, and a leg controller, enabling the synthesis of a wide range of motions.

Datasets

Not discussed in the paper. The evaluation is performed on hardware platforms.

Key Insights and Ideas

  • Unified Control for Agility: The most significant insight is that a single MPC framework, using a sophisticated hybrid model and solver, can effectively generate both regular gaits and highly dynamic, aperiodic motions like jumping. This breaks down the traditional separation of controllers.
  • The Power of Feedback in MPC: Incorporating the DDP feedback gain into the MPC's initial roll-out dramatically improves robustness and convergence, especially in the face of state discontinuities like leg touchdowns. This is a crucial detail for real-time applications.
  • Efficient DDP for Hybrid Systems: The HS-DDP solver, with its specific schemes for handling state resets and re-initialization, is key to making complex kinodynamic MPC computationally feasible for real-time control on hardware.
  • Heuristic References are Sufficient: The paper demonstrates that even simple, human-authored heuristic references (e.g., desired jump height and speed) can lead to successful and robust complex maneuvers when paired with a powerful controller like HKD-MPC. This suggests that the controller itself is capable of inferring the necessary dynamics.

Key Results

  • Versatile Motion Synthesis: The HKD-MPC successfully generated and transitioned between diverse motions, including:
  • Run-jump-run sequences (jumping up to 0.5m high and 1.0m far on Unitree A1).
  • Mixed-gaits hopping (alternating diagonal legs, four legs, and trotting).
  • Continuous jumping in place.
  • Hardware Performance: The algorithm was successfully implemented and demonstrated on two quadruped platforms: Unitree A1 and MIT Mini Cheetah.
  • Real-time Operation: The MPC controller runs at 100 Hz, with planning horizons of 0.5 seconds and integration time steps of 0.01 seconds.
  • Fast Solve Times: Table I shows average solve times for tasks like "Run-jump-run" at around 5.64 ms, enabling the 100 Hz control loop.
  • Transferability: Switching between robot platforms (Unitree A1 and Mini Cheetah) only required changing mass, inertia, and kinematic parameters, highlighting the controller's generality.

Limitations and Future Work

  • Human-Authored References: While the controller works well with simple heuristic references, the authors acknowledge that more realistic references, typically generated by trajectory optimization, could lead to even better performance. However, generating these references can be computationally slow.
  • Future Work: The authors suggest investigating the use of machine learning for generating more realistic and complex reference trajectories. This would combine the strengths of learned motion generation with the robustness of MPC for control.

Summary updated

Abstract

Specialized motions such as jumping are often achieved on quadruped robots by solving a trajectory optimization problem once and executing the trajectory using a tracking controller. This approach is in parallel with Model Predictive Control (MPC) strategies that commonly control regular gaits via online re-planning. In this work, we present a nonlinear MPC (NMPC) technique that unlocks on-the-fly re-planning of specialized motion skills and regular locomotion within a unified framework. The NMPC reasons about a hybrid kinodynamic model, and is solved using a variant of a constrained Differential Dynamic Programming (DDP) solver. The proposed NMPC enables the robot to perform a variety of agile skills like jumping, bounding, and trotting, and the rapid transition between these skills. We evaluated the proposed algorithm with three challenging motion sequences that combine multiple agile skills, on two quadruped platforms, Unitree A1, and MIT Mini Cheetah, showing its effectiveness and generality.
BibTeX
@inproceedings{yu2023versatile,
  author = {Wenhao Yu and He Li and Tingnan Zhang and Wenhao Yu and Patrick M. Wensing},
  title = {{Versatile Real-Time Motion Synthesis via Kino-Dynamic MPC With Hybrid-Systems DDP}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  doi = {10.1109/icra48891.2023.10160221},
}
Figure