Global Position Control on Underactuated Bipedal Robots: Step-to-step Dynamics Approximation for Step Planning

Xiaobin Xiong, Jenna Reher, Aaron D. AmesICRA 2021 · DOI

TL;DR

This paper presents a novel framework for enabling underactuated bipedal robots to achieve global position control, allowing them to follow arbitrary paths. By approximating the robot's complex dynamics with a simplified Hybrid Linear Inverted Pendulum (H-LIP) model and employing Model Predictive Control (MPC), the system can plan and execute stable walking gaits for precise navigation.

What Problem Does This Paper Address?

Controlling the global position of bipedal robots, meaning making them follow a specific path or reach a target location, is a well-established problem for fully actuated robots. However, for underactuated bipedal robots (robots where not all joints are independently controllable, particularly the feet), this task is significantly more challenging. The lack of direct actuation on the feet makes balancing and precise trajectory following difficult, as the robot's motion is heavily reliant on its natural dynamics and careful foot placement. Existing research has largely focused on generating stable, local walking behaviors or periodic gaits, rather than enabling global path following in complex environments. This paper aims to bridge this gap by developing a method for underactuated bipeds to achieve robust global position control.

Methods

The core of the proposed method lies in approximating the complex, full-order dynamics of the underactuated bipedal robot with a simplified model called the Hybrid Linear Inverted Pendulum (H-LIP). The H-LIP captures the essential step-to-step (S2S) dynamics of the robot's horizontal motion.

Here's a breakdown of the key techniques:

  • H-LIP Approximation: The paper leverages the H-LIP model, which is a passive model in its single support phase (SSP) and has simplified dynamics in its double support phase (DSP). Crucially, the S2S dynamics of the H-LIP are used to approximate the S2S dynamics of the actual robot. This approximation treats the differences between the H-LIP and the real robot as disturbances.
  • Model Predictive Control (MPC) on H-LIP: MPC is applied to the H-LIP model to plan future step sizes. The MPC optimizes a cost function that balances tracking a desired trajectory with penalizing large step sizes, ensuring smooth and feasible motions.
  • Feedback Adjustment: A feedback controller is designed to ensure that the actual robot's state stays close to the H-LIP's state. This is achieved by using the H-LIP's planned step sizes and adjusting them based on the difference between the robot's and the H-LIP's states. This feedback mechanism helps bound the error introduced by the H-LIP approximation.
  • 3D Extension: The planar H-LIP model is extended to 3D by treating the sagittal and lateral planes independently. A trivial torso is added to the H-LIP to define orientation, enabling turning capabilities.
  • Output Construction for Full-Order Robot: The planned step sizes from the H-LIP are then translated into specific control commands for the actual robot. This involves generating periodic leg length trajectories, defining swing foot trajectories, and controlling the pelvis and swing foot orientations.
  • Low-Level Feedback Control: A Control Lyapunov Function (CLF) based quadratic program (CLF-QP) is used as a low-level controller to stabilize the robot's outputs (e.g., desired leg length, foot position) to zero, ensuring the robot accurately follows the generated trajectories.

Datasets

Not discussed in the paper. The paper focuses on simulation and experimental verification of the proposed control framework on a specific robot model (Cassie).

Key Insights and Ideas

  • Approximation as a Powerful Tool: The paper demonstrates that a simplified model (H-LIP) can be effectively used to control a much more complex system (underactuated biped). The key insight is to treat the model mismatch as a bounded disturbance, which can be managed by feedback.
  • Decoupling Global Control and Local Stability: The framework elegantly separates the problem into two parts: global path planning using the H-LIP and MPC, and local stability and trajectory tracking using the full-order robot dynamics and CLF-QP. This modular approach simplifies the design.
  • Turning as an Integrated Feature: By incorporating a trivial torso and extending to 3D, the H-LIP framework naturally supports turning, which is crucial for robots navigating in real-world environments and aligning their sensors.
  • Dynamic Consistency for Hardware Transfer: The paper highlights the importance of ensuring that the motion generated in simulation is dynamically achievable on the actual hardware. This is demonstrated by the experimental validation on Cassie.

Key Results

  • Successful Path Tracking: The proposed method was successfully demonstrated in simulation for tracking various paths, including circles, cardioids, sinusoids, and squares, even at relatively high speeds.
  • Obstacle Avoidance and Disturbance Rejection: The framework was extended to handle challenging scenarios, including avoiding obstacles by modifying the MPC constraints (using QCQP) and rejecting external disturbances (e.g., a 200N lateral force).
  • Experimental Validation: A proof-of-concept experiment was conducted on the Cassie robot, showing that the simulated walking trajectories are dynamically achievable on hardware. The experimental results closely matched the simulation, validating the approach.
  • Error Bounding: The tracking errors were shown to remain within the predicted disturbance invariant set, confirming the theoretical underpinnings of the method.

Limitations and Future Work

  • Approximation Error: The paper acknowledges that the H-LIP is an approximation, and the error between the H-LIP and the full robot dynamics is treated as a disturbance. Future work aims to improve this approximation through numerical methods to reduce the disturbance term.
  • Hardware Realization: While a proof-of-concept experiment was performed, the paper states that a full realization of the approach on hardware, including global position feedback, is a key area for future work.
  • Robustness to Larger Disturbances/Maneuvers: The paper notes that dynamic maneuvers like sharp turning at high speeds can be challenging and may lead to tracking failure if step sizes exceed motor torque limits. Future work could explore more robust controllers to minimize the disturbance invariant set and improve performance in such scenarios.
  • Systematic Output Construction: Improving the systematic construction of outputs for the full-order robot to make it closer to the H-LIP walking is also identified as a future direction.

Summary updated

BibTeX
@inproceedings{xiong2021global,
  author = {Xiaobin Xiong and Jenna Reher and Aaron D. Ames},
  title = {{Global Position Control on Underactuated Bipedal Robots: Step-to-step Dynamics Approximation for Step Planning}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2021},
  doi = {10.1109/icra48506.2021.9561961},
}
Figure