TL;DR
This paper introduces a new framework for quadruped robots to recover from pushes by characterizing "dynamic balance" and using this to plan footstep adjustments. This approach significantly improves disturbance rejection compared to existing methods, allowing robots to handle larger impacts and adapt to different gaits and push timings.
What Problem Does This Paper Address?
Quadruped robots are excellent for navigating challenging terrain, but they are also prone to falling when pushed. Existing methods for push recovery, largely developed for bipeds, often rely on achieving static or quasi-static balance, which is too restrictive for dynamic quadrupedal locomotion. Quadrupedal robots have unique features like multiple swinging legs and point contacts, and their balance is inherently time-varying due to their gaits. This paper addresses the challenge of developing robust push recovery strategies that account for these quadruped-specific dynamics, the time-varying nature of balance, and the influence of gait and impact timing.
Methods
The authors propose a novel approach based on switched systems modeling to capture the dynamics of quadruped locomotion. They extend the concept of capturability by introducing dynamic balance, which allows for controlled movement during recovery rather than requiring an immediate stop. This dynamic balance is formalized as a tube of dynamically balanced states.
To analyze this, they employ Explicit Model Predictive Control (EMPC) to compute these "balanced" and "capturable" tubes. This offline analysis provides a detailed understanding of how gait patterns and disturbance timings affect the robot's ability to recover. Based on this, they develop an online CoM and footstep planning scheme. This scheme iteratively plans the robot's center of mass (CoM) trajectory and footstep locations to ensure the robot lands in a capturable state, using quadratic programming for efficient real-time execution.
Datasets
Not discussed in the paper. The authors use simulations and hardware experiments on a Mini Cheetah robot.
Key Insights and Ideas
- Dynamic Balance is Crucial: Static balance is insufficient for dynamic quadrupedal locomotion. The paper formalizes "dynamic balance" as a controlled evolution of the robot's state within a time-varying tube, allowing for more flexible recovery.
- Gait and Timing Matter: The ability to recover from a push is highly dependent on the robot's current gait and the timing of the impact. The proposed framework explicitly accounts for these factors, leading to more robust performance.
- Capturability as a Tube: Instead of a single point, capturability is characterized by a "tube" of states that can lead to dynamic balance, providing a richer understanding of recovery capabilities.
- EMPC for Analysis: EMPC is used effectively to analyze the complex dynamics and derive the necessary sets for capturability analysis, providing valuable offline insights.
- Online Planning for Real-time Recovery: The framework translates offline analysis into an efficient online planning strategy that adjusts footstep placement and CoM trajectories to achieve recovery.
Key Results
- Significant Improvement: The proposed approach demonstrates an improvement of up to 100% in rejecting external pushes compared to state-of-the-art methods (specifically, a convex MPC and whole-body impulse control approach).
- Enhanced Disturbance Rejection: Simulations show the proposed method can reject impacts up to twice as large as the baseline, and in some directions, up to three times larger.
- Hardware Validation: Experiments on a Mini Cheetah robot confirm the effectiveness of the proposed strategy, showing improved performance and robustness.
- Gait and Timing Influence: The results clearly illustrate how different gaits and impact timings affect the robot's recovery capabilities, a factor often overlooked in previous methods.
Limitations and Future Work
- Simulation vs. Hardware Discrepancy: While simulations show high disturbance rejection, hardware experiments exhibit a larger fluctuation in balance due to factors like state estimation noise and unmodeled dynamics.
- State Estimation: The performance in hardware is sensitive to the accuracy of the state estimator. Better-tuned parameters or more advanced estimators could further improve performance.
- Self-Collision and Trajectory Tracking: Issues like self-collision and imperfect swing leg trajectory tracking can lead to recovery failures in hardware, which are not modeled in simulations.
- Future Work: The authors suggest investigating gait adjustment strategies in reaction to disturbances as a future research direction. They also highlight the need for better state estimation and addressing hardware-specific issues for practical deployment.
Summary updated