TL;DR
This paper demonstrates that complex humanoid tasks like balancing and walking can be achieved using a simple set of contact-consistent LQR controllers instead of computationally heavy per-cycle optimization. The authors show that linearizing the robot's full dynamics at just a few key contact configurations is sufficient for robust performance on a real hydraulic humanoid, rivaling the performance of far more complex state-of-the-art controllers.
What Problem Does This Paper Address?
Humanoid control is notoriously difficult due to high degrees of freedom, non-linear dynamics, and the "hybrid" nature of walking, where contact states (feet on or off the ground) constantly switch. The current research trend favors high-complexity strategies, such as solving Quadratic Programs (QPs) at every millisecond to optimize joint torques.
While effective, these methods carry significant computational overhead and lack the clear stability metrics (like gain and phase margins) found in classical control. The authors ask a fundamental question: How much complexity is actually required? They explore whether a simpler, "old-school" Linear Quadratic Regulator (LQR) approach—if properly adapted for contact constraints—can handle the dynamic balancing and walking tasks typically reserved for more complex optimization-based frameworks.
Methods
The authors extend a "full dynamics" LQR framework to handle changing environments:
- Contact-Consistent Projection: Standard LQR doesn't know about floor constraints. The authors project the linearized robot dynamics into the nullspace of kinematic contact constraints. This ensures the controller respects that a foot in contact should have zero velocity and acceleration. Crucially, their new projection method is valid for non-static poses, allowing for linearization along moving trajectories.
- Key-Pose Linearization: Instead of re-calculating the optimal control law at every time step, they pre-calculate LQR gains for a small set of "key poses" (e.g., standing centered, leaning left, leaning right).
- Gain Scheduling and Blending: The robot switches between these pre-calculated gain sets based on its estimated contact state. To prevent jerky torque jumps during transitions, they use a heuristic blending method to smoothly interpolate between controllers.
Datasets
Not discussed in the paper. The research relies on real-time sensor data from a physical hydraulic Sarcos humanoid robot and physics-based simulations.
Key Insights and Ideas
- Contacts Matter More Than Poses: A major finding is that optimal control gains change drastically when the contact state changes (e.g., lifting a foot), but they change very little as the robot moves within the same contact state.
- Diminishing Returns of Complexity: On the real robot, increasing the number of linearizations along a path did not improve tracking performance. The authors argue that hardware-level bottlenecks—like sensor noise and torque tracking limits—dominate the system's performance long before control law complexity does.
- The Power of Coupling: Unlike standard independent joint controllers (PD control), LQR explicitly captures the "synergies" or couplings between joints. For example, it understands how moving a hip joint affects the balance of the entire base. This coupling is what allows a "simple" linear controller to perform complex whole-body coordination.
Key Results
The framework was tested on a lower-body hydraulic Sarcos humanoid:
- Internal Disturbance Rejection: The robot successfully balanced while a 10kg mass on its torso moved sinusoidally at up to 0.8 Hz in the sagittal plane.
- External Push Recovery: In single-leg support, the controller rejected external impulses of 8.2 Ns (with peak forces of 488 N). This is competitive with results from much more complex hierarchical QP controllers (which achieved ~5.8 Ns on the same hardware).
- Walking: The robot achieved stable walking in both simulation and reality using only three sets of LQR gains (one for double support, and one for each single-support pose).
Limitations and Future Work
- Noise Sensitivity: LQR tends to produce aggressive velocity feedback. On real hardware, this can amplify sensor noise, requiring heavy filtering that can limit the overall control bandwidth.
- Torque Control Dependency: The high-level LQR performance is strictly limited by the quality of the low-level torque control loop. The authors found that overly aggressive low-level compensation could actually destabilize the LQR controller.
- Open-Loop Trajectory Limits: The walking experiments relied on pre-planned trajectories. If the robot's foot placement deviates significantly, the pre-planned path can become kinematically impossible. The authors suggest that integrating this LQR approach with online model predictive control (MPC) planners would be a powerful next step for more robust, adaptive walking.
Summary updated