TL;DR
This paper introduces a novel force-feedback based whole-body stabilizer for position-controlled humanoid robots. By fully leveraging six-dimensional contact force measurements and the robot's complete dynamics, it significantly improves tracking performance and enables robots to handle more challenging contact scenarios compared to existing methods.
What Problem Does This Paper Address?
Humanoid robots, especially those controlled by position commands (common in robots like Honda ASIMO or UBTECH Walker), face significant challenges in maintaining stable locomotion. These challenges stem from complex whole-body dynamics, underactuated floating bases, intricate contact behaviors, and the inherent limitations of position-controlled actuators. Existing control strategies often rely on simplified models like the Linear Inverted Pendulum (LIP) and Zero-Moment Point (ZMP) feedback. While these approaches have achieved some success, they often fall short in complex or dynamic environments because they don't fully utilize the available information, particularly the rich data from six-dimensional force sensors, and don't comprehensively consider the robot's entire body dynamics. This paper aims to bridge this gap by developing a stabilizer that integrates these crucial elements for more robust and adaptable control.
Methods
The core of this paper is a novel force-feedback based whole-body stabilizer. The approach begins with a rigorous analysis of the whole-body dynamics of position-controlled humanoids, specifically addressing the complexities introduced by unknown contact conditions. This analysis leads to a fundamental relationship between joint position commands and the robot's Center of Mass (CoM) behavior and actual contact forces.
Leveraging this relationship, the stabilizer design is formulated as an optimization problem solved using quadratic programming (QP). This QP formulation allows for the cooperative consideration of two key tasks:
- Center-of-Mass (CoM) tracking: Ensuring the robot's overall balance and trajectory.
- Contact force tracking: Precisely controlling the forces exerted at the robot's feet.
The stabilizer employs a task-space differential inverse kinematics (Differential IK) approach. This method computes joint velocities first, which are then integrated to obtain joint position commands. This strategy is advantageous because it linearizes the relationship between joint and end-effector velocities, making the QP formulation tractable. The process involves two main QP steps:
- Inverse Dynamics QP: Generates a desired contact force that respects dynamic constraints and tracks the desired CoM.
- Differential IK QP: Computes the optimal joint velocities to simultaneously achieve CoM and contact force tracking, while also minimizing effort and ensuring closed-loop error dynamics for contact forces.
Datasets
Not discussed in the paper. The evaluation is performed using simulations on the UBTECH Walker robot within the MuJoCo simulator.
Key Insights and Ideas
- The Power of Full Six-Dimensional Force Feedback: The paper strongly emphasizes that utilizing the complete six-dimensional force measurement (not just ZMP) is crucial for significantly improving stabilization performance, especially in dynamic and uncertain environments.
- Whole-Body Dynamics are Essential: Moving beyond simplified models like LIP and ZMP, this work demonstrates the benefit of incorporating the robot's entire dynamic model into the stabilizer design. This allows for a more holistic approach to control.
- Cooperative Optimization of CoM and Contact Forces: By formulating the problem as a single QP, the stabilizer can jointly optimize for both CoM stability and desired contact forces, revealing synergistic benefits that are missed by decoupled approaches.
- Differential IK for Tractability: The use of differential inverse kinematics transforms a potentially complex nonlinear problem into a more manageable quadratic programming problem, making real-time implementation feasible.
- Admittance Control with Adaptive Gain: The derived control law for force error resembles an admittance control law, but with a gain matrix that is adaptively related to the robot's contact inertia, offering a more sophisticated form of compliance.
Key Results
The paper demonstrates significant improvements in tracking performance and robustness through simulations on the UBTECH Walker robot in MuJoCo.
- Improved Tracking: The proposed stabilizer shows a significant improvement in various scenarios compared to commonly adopted ZMP-based stabilizers.
- Enhanced Robustness: The stabilizer is capable of handling more challenging scenarios, including continuously changing platforms and unexpected discontinuous terrain changes, where classical ZMP-based strategies fail.
- Reduced Variations: In scenarios with changing platforms, the proposed stabilizer achieved much smaller CoM variations and smaller contact forces compared to the baseline ZMP feedback approach.
- Successful Recovery: In a critical scenario where a support platform abruptly declines, the proposed stabilizer successfully balances the robot and retains contact, while the ZMP-based stabilizer fails.
Limitations and Future Work
- Actuator Bandwidth Limitations: The authors acknowledge that the proposed stabilizer's performance is limited in situations with very abrupt terrain changes (larger than approximately 60 deg/s roll/pitch rate of change). This limitation is attributed to the limited bandwidth of the position-controlled actuators.
- Future Work: Potential extensions include:
- Integrating the proposed stabilizer with high-level planners to achieve robust dynamic locomotion (e.g., walking and running).
- Further validations on hardware platforms.
Summary updated