TL;DR
This paper presents a novel hierarchical control architecture for humanoid robots that enables robust, compliant walking by combining online predictive planning for footstep and CoM trajectories with a stack-of-tasks torque controller. This approach allows for dynamic adjustments to walking paths and improved stability, demonstrated on the iCub robot.
What Problem Does This Paper Address?
Bipedal locomotion in humanoid robots remains a significant challenge due to unpredictable terrain, complex robot-environment interactions, and the inherent limitations of robotic actuators compared to biological systems. Existing approaches often rely on complex models that are computationally expensive or require extensive tuning. This paper aims to develop a more robust and adaptable control architecture that can handle dynamic changes in desired walking paths and provide greater compliance, particularly in torque-controlled walking scenarios.
Methods
The proposed architecture is built upon three nested control loops:
- Outer Loop (Footstep Planner): This module uses a simplified unicycle model to plan footstep positions. It takes into account the robot's current state and a desired direction from a user (e.g., a joystick) to generate a sequence of foot placements. Unlike some prior work, it doesn't fix the robot's velocity and can handle situations where the robot is not in single support.
- Mid Layer (Receding Horizon Control - RHC): This layer generates kinematically feasible trajectories for the robot's Center of Mass (CoM) and joint positions. It utilizes a Linear Inverted Pendulum (LIP) model and whole-body inverse kinematics. The RHC module operates with time-varying constraints to keep the Zero Moment Point (ZMP) within the support polygon, enhancing robustness.
- Inner Loop (Stack-of-Tasks Balancing Controller): This is a Quadratic Programming (QP)-based torque controller that takes the CoM and foot trajectories from the RHC as desired values. It computes the necessary joint torques to achieve these trajectories while satisfying various tasks (e.g., CoM stabilization, foot placement, joint position tracking) and constraints (e.g., friction, contact forces). This layer is crucial for enabling torque control and achieving compliance.
The architecture is designed to be flexible, allowing it to be implemented on robots controlled either by position or torque.
Datasets
Not discussed in the paper. The experiments were conducted on the iCub humanoid robot.
Key Insights and Ideas
- Hierarchical Control for Adaptability: The three-layered approach effectively decomposes the complex problem of walking into manageable sub-problems, allowing for online adaptation of footstep and CoM trajectories based on user input or environmental changes.
- Unicycle Model for Footstep Planning: Using a simplified unicycle model for footstep planning, while seemingly basic, proves effective for generating dynamic walking paths and is computationally efficient.
- RHC with Time-Varying Constraints: Incorporating time-varying constraints into the RHC for ZMP stabilization significantly improves robustness by preventing overshoots and potential tipping.
- Stack-of-Tasks for Compliance: The stack-of-tasks QP controller is key to achieving compliant walking. By formulating walking as a set of prioritized tasks and constraints, it can generate joint torques that allow the robot to absorb external disturbances and interact more safely.
- Bridging Position and Torque Control: The architecture's ability to seamlessly transition between position and torque control is a significant advantage, enabling more robust and compliant locomotion.
Key Results
The paper demonstrates the effectiveness of the proposed architecture through experiments on the iCub humanoid robot.
- Position Control: The robot successfully walks in position control, tracking desired trajectories.
- Online Path Adaptation: With the addition of the unicycle planner, the robot can adapt its walking direction online (e.g., turning right) based on joystick input. Step timings and dimensions are no longer fixed, allowing for more natural gait variations.
- Torque Control and Compliance: The full architecture, including the stack-of-tasks balancing controller, enables torque-controlled walking. This results in smoother motion and increased robustness to disturbances, as evidenced by improved CoM tracking and reduced likelihood of falling. The minimum stepping time was increased to 3 seconds for better stability in torque control.
Limitations and Future Work
- Joint Torque Estimation: The authors acknowledge that the estimation of joint torques is a bottleneck that currently limits the performance of the torque-controlled walking.
- Floating Base Estimation: Similar to joint torque estimation, the accuracy of the floating base estimation also impacts performance.
- Future Work: The authors plan to address these limitations by increasing the robustness of the architecture to uncertainties in joint torque and floating base estimation. They also aim to further explore the potential of compliant torque-controlled walking for handling un-modeled ground slope variations.
Summary updated