Introduction
In classical robotics, the "separation principle" typically dictates that state estimation and control are treated as independent modules. A standard Extended Kalman Filter (EKF) estimates the most likely state, which a controller then uses to determine actions. However, this approach ignores the context of the task. For instance, a small estimation error in a robot's height is negligible on flat ground but critical when navigating obstacles. Standard filters often suffer from "optimism bias," assuming the most probable state is safe until a failure occurs.
This research introduces the Risk-Sensitive Extended Kalman Filter (RS-EKF), a state estimator that incorporates the robot's control objectives into its estimation process. By adopting a "cautiously pessimistic" stance, the RS-EKF biases its estimates toward states that would be detrimental to the task, thereby forcing the controller to act more robustly.
Methodology
The RS-EKF is built upon the framework of Dynamic Game Output Feedback Model Predictive Control (MPC). It frames estimation as a zero-sum game where an imaginary adversary selects disturbances to maximize the robot's task cost.
The Risk-Sensitive Shift
The core innovation lies in integrating the Value Function—the predicted future cost—from the MPC controller into the filter's update rule. The RS-EKF estimate is derived as the standard EKF estimate plus a bias term proportional to the estimation uncertainty ($P_t$) and the gradient of the value function.
When uncertainty is low, the RS-EKF behaves like a standard EKF. However, when uncertainty increases, the filter shifts its estimate toward "risky" regions of the state space. This coupling ensures that the estimator is "aware" of which errors matter most for the current goal. Crucially, this is achieved with negligible computational overhead, making it suitable for high-frequency real-time control loops (e.g., 500Hz).
Experimental Validation
The performance of the RS-EKF was evaluated across three distinct scenarios:
- Planar Quadrotor (Simulation): A drone carrying an unknown 3kg load that is suddenly released.
- Kuka LWR iiwa (Simulation): An industrial manipulator subjected to 10,000 trials of random external disturbances during trajectory tracking.
- Solo12 Quadruped (Hardware): A real-world torque-controlled robot tested for balance and external force estimation.
Results and Discussion
Quadrotor Mass Estimation
In the quadrotor simulation, the RS-EKF demonstrated significantly higher reactivity to sudden changes in mass compared to a standard EKF. By being "pessimistic" about the load, the filter allowed the controller to compensate more effectively, leading to a 54% improvement in tracking accuracy (MSE) and a 35% reduction in total cost.
Solo12 Hardware Performance
On the Solo12 quadruped, the RS-EKF's task-aware bias proved vital for maintaining stability under external forces. When a 20N pulling force was suddenly released, the RS-EKF "underestimated" the robot's height. This pessimism prompted the MPC to generate higher ground reaction forces, returning the robot to its target posture significantly faster than the standard EKF. This behavior resulted in a 50-60% improvement in average cost on real hardware.
Conclusion and Limitations
The RS-EKF offers a computationally efficient way to bridge the gap between estimation and control. By introducing uncertainty-triggered robustness, it enables robots to handle unexpected disturbances more gracefully.
However, the "pessimistic" bias can introduce small steady-state errors, as the robot may constantly hedge against non-existent worst-case scenarios. Additionally, the risk-sensitivity parameter ($\mu$) requires careful tuning to prevent the underlying optimization from becoming ill-posed. Future work will explore applying this framework to highly dynamic maneuvers like jumping and more complex noise models.
Summary updated