Author: Vu Nhat Minh (KIST/UST)
Abstract
Controlling humanoid robots involves managing high-dimensional degrees of freedom (DoFs), which often leads to significant computational overhead. This research proposes a hierarchical control strategy that utilizes a simplified "template" model—the Bipedal Trunk Spring-Loaded Inverted Pendulum (BTSLIP)—to guide complex, multi-link robots. By introducing a novel Force Direction Controller (FDC), the study demonstrates that robots can maintain stable upright postures and navigate rough terrain or external disturbances more reliably than previous state-of-the-art methods like the Virtual Pivot Point (VPP).
Introduction
Humanoid locomotion is a complex challenge due to the high number of actuators and the inherent instability of bipedal structures. To simplify this, researchers often employ "reduced-order models"—physics-based skeletons that capture essential dynamics without the computational cost of full-body modeling.
A common approach is the Virtual Pivot Point (VPP) method, which redirects ground reaction forces (GRF) toward a point above the robot's center of mass to maintain balance. However, this research identifies a critical limitation: in specific body configurations, the VPP method can generate an "upsetting moment" that destabilizes the robot. This paper seeks to address these vulnerabilities through a more robust, hierarchical control architecture.
Methodology
The proposed control strategy is divided into two distinct layers:
Layer 1: The Planner (BTSLIP Model)
The high-level planner uses the Bipedal Trunk Spring-Loaded Inverted Pendulum (BTSLIP) model, which treats the robot's legs as virtual springs and the body as a rigid mass with a trunk.
The core innovation here is the Force Direction Controller (FDC). Unlike the static target of the VPP, the FDC uses a PD-like control law to dynamically adjust the angle of the GRF based on the trunk's instantaneous pitch and velocity. This ensures that the force always provides a "righting moment," effectively keeping the torso vertical regardless of the robot's configuration.
Layer 2: The Executor
This layer translates abstract forces from the BTSLIP model into specific joint torques for articulated robots:
- 5-Link Model: Utilizes Operational Space Control (OSC) and Jacobian Transpose methods to map desired foot forces to joint movements.
- 7-Link Model: Incorporates ankles and flat feet, managed by a Finite State Machine (FSM) to handle transitions such as heel-strike and ankle push-off.
Simulation and Results
The framework was validated using high-fidelity physics simulations in MATLAB and the Open Dynamics Engine (ODE). Key findings include:
- Disturbance Rejection: The controller successfully recovered from external force impulses of up to 300N, demonstrating superior stability compared to traditional methods.
- Blind Terrain Adaptation: The 7-link robot successfully traversed rough terrain with height variations up to 5cm (approximately 20% of leg length) without prior knowledge of the ground geometry.
- Emergent Dynamics: The robot's walking speed stabilized naturally through the interaction of springy leg dynamics and the Velocity-Based Leg Adjustment (VBLA) algorithm. Furthermore, the system naturally produced the "double-peak" GRF pattern characteristic of human gait.
- Computational Efficiency: The FDC achieved stability comparable to complex optimization-based methods (like DLQR) but with significantly lower overhead, making it ideal for real-time hardware applications.
Conclusion and Future Work
This hierarchical approach provides a computationally efficient and robust solution for planar bipedal walking. By focusing on the direction of ground reaction forces rather than a fixed virtual point, the system achieves high levels of adaptability to unknown terrain.
Future research will focus on extending this state-based control to 3D humanoid models to address lateral balance and implementing the strategy on physical hardware to account for real-world sensor noise and motor limitations.
Summary updated