TL;DR
This paper introduces a novel framework for humanoid robots that allows them to dynamically recover from pushes while walking, leveraging their arms to brace against environmental features like walls. By combining a Single Rigid Body Model Predictive Control (SRB-MPC) with a Hybrid Linear Inverted Pendulum (HLIP) model, the robot achieves significantly improved robustness and stability compared to existing methods, enabling it to maintain locomotion even under substantial external forces.
What Problem Does This Paper Address?
Humanoid robots are envisioned for complex, human-centered environments, but their inherent instability makes them vulnerable to external perturbations like pushes. While existing research has focused on either robust locomotion controllers or specific push recovery modes, a unified approach that seamlessly integrates both, especially by utilizing the robot's full body capabilities, has been lacking. Specifically, the ability for a humanoid robot to dynamically use its arms to brace against its surroundings (like walls) for recovery while simultaneously maintaining its walking gait is a critical but challenging problem that this paper aims to solve. This is crucial for real-world deployment where unexpected interactions are common.
Methods
The core of the proposed solution is a unified framework that combines two reduced-order models (ROMs) within a Model Predictive Control (MPC) approach:
- Single Rigid Body (SRB) Model: This model simplifies the robot's dynamics by treating it as a single rigid body. It's extended to handle multiple contact points (hands, feet) and is used to generate a generalizable model for control.
- Hybrid Linear Inverted Pendulum (HLIP) Model: This model is used to generate robust support polygons and stepping strategies. It's particularly effective at handling perturbed Center of Mass (CoM) velocities and adjusting step sizes for stability.
The framework integrates these two models into an SRB-MPC-HLIP controller. When a push is detected (by observing deviations from nominal states), the controller activates a "recovery mode." This mode leverages the robot's arms to brace against environmental features (like walls) by dynamically adjusting contact forces and support polygons. The HLIP model helps in generating appropriate stepping patterns, while the SRB-MPC handles the overall balance and control, including the forces exerted by the arms. The system also incorporates environment sensing (e.g., wall detection via LiDAR) and state estimation.
Datasets
Not discussed in the paper. The evaluation is performed in a high-fidelity simulation environment (MuJoCo [23]).
Key Insights and Ideas
- Leveraging the Environment for Recovery: The most significant insight is the effective use of environmental features (walls) as a recovery mechanism. This moves beyond solely relying on the robot's internal dynamics and opens up new possibilities for robust interaction.
- Synergy of SRB and HLIP Models: Combining the SRB model's ability to handle multi-contact forces with the HLIP model's robust stepping and support polygon generation creates a powerful synergy for both locomotion and recovery.
- Unified Framework: The paper successfully demonstrates a unified approach that seamlessly integrates push detection, environmental interaction, and dynamic locomotion, rather than treating them as separate modes.
- Dynamic Adjustment of Constraints: The controller dynamically adjusts contact constraints (e.g., foot and hand forces) based on the current contact states and MPC projections, allowing for more adaptive recovery.
Key Results
- Improved Perturbation Rejection: The SRB-MPC-HLIP controller significantly outperforms a pure HLIP controller in rejecting external forces. In worst-case scenarios without environmental assistance, the proposed controller can reject up to 30N in all directions, a 50% increase over the pure HLIP controller (20N).
- Larger "Safeset": When tested with a range of push forces (30-100N) and heights, the SRB-MPC-HLIP controller successfully recovered in 70.6% of cases, compared to only 16.8% for the nominal HLIP controller. This represents a 420% increase in the "safeset" size.
- Robustness to Environment and Push Variations: The controller demonstrated robustness in scenarios with angled walls and multi-directional pushes, successfully recovering and continuing to walk.
- Tracking Performance: The proposed controller maintains adequate command tracking performance while also improving torso orientation stabilization compared to the pure HLIP controller.
Limitations and Future Work
- Simulation-Based: The current evaluation is performed in a high-fidelity simulation. Deploying the framework on hardware is a crucial next step.
- External Disturbances Only: The current work focuses on external pushes. Future work could extend the framework to include internal disturbances caused by the robot's own limb movements.
- Learned Push Prediction: The authors suggest using a learned classifier to detect and predict pushes, which could further enhance the system's proactive capabilities.
- Generalizability: Adapting the method to different humanoid platforms and varying environmental surfaces and objects is a key area for future research.
Summary updated