TL;DR
This paper introduces the Risk-Sensitive Extended Kalman Filter (RS-EKF), a state estimator that biases its predictions based on the robot's control goals. By being "pessimistic" in the face of uncertainty—assuming the state might be in a region that makes the task harder—it forces the controller to act more robustly, significantly improving performance on drones and legged robots without increasing computational cost.
What Problem Does This Paper Address?
In classical robotics, we usually follow the separation principle: one module estimates the robot's state (where it is, how fast it's moving) using a tool like an Extended Kalman Filter (EKF), and another module decides what to do (the controller).
The problem is that standard EKFs only care about the "most likely" state. They don't know why the estimate matters. For a quadruped robot, an error of 2cm in height is trivial if it's standing on flat ground, but catastrophic if it's trying to clear a step. When uncertainty is high—like when a robot picks up an unknown heavy load—relying on the "average" estimate can lead to "optimism bias," where the robot assumes everything is fine until it suddenly isn't.
The authors aim to bridge this gap by creating a filter that is "aware" of the control objective, allowing for safer output-feedback Model Predictive Control (MPC) in nonlinear systems.
Methods
The authors leverage a framework called Dynamic Game Output Feedback MPC. They treat the estimation process as a zero-sum game: the robot wants to minimize its task cost, while an imaginary "adversary" chooses disturbances to maximize that cost.
Key Innovations:
- Value Function Integration: Unlike a standard EKF, the RS-EKF uses the Value Function (the predicted future cost) from the MPC controller. Specifically, it uses the quadratic approximation of the value function provided by algorithms like Differential Dynamic Programming (DDP).
- The Risk-Sensitive Shift: The core update rule (Equation 17 in the paper) shows that the RS-EKF estimate is the standard EKF estimate plus a "bias" term. This bias is proportional to the estimation uncertainty ($P_t$) and the gradient of the value function.
- Computational Parity: Remarkably, the RS-EKF has the same computational complexity as a standard EKF. It essentially adds a single linear algebra step to the update, making it perfectly suited for high-frequency, real-time control loops (e.g., 500Hz).
Datasets
As a robotics control paper, the evaluation is performed across three distinct experimental setups rather than static datasets:
- Planar Quadrotor (Sim): A drone tasked with carrying an unknown 3kg load that is suddenly dropped.
- Kuka LWR iiwa (Sim): A 7-DOF industrial manipulator performing trajectory tracking while being subjected to 10,000 trials of random external "pushes."
- Solo12 Quadruped (Real Hardware): A torque-controlled quadruped robot used to validate the filter's performance in estimating external forces and maintaining balance in the real world.
Key Insights and Ideas
- Task-Aware Pessimism: The filter doesn't just guess the state; it guesses a state that would be "bad" for the current task. If the robot is near a cliff, the filter will bias the estimate toward the cliff, forcing the controller to steer away sooner.
- Uncertainty-Triggered Robustness: The "pessimistic shift" is scaled by the filter's covariance ($P_t$). When the filter is very certain, it behaves like a normal EKF. When uncertainty spikes (e.g., a sudden impact), the risk-sensitivity kicks in, making the robot act more cautiously.
- Closing the Loop: This creates a tight coupling between estimation and control. The controller provides the "risk map" (value function), and the estimator uses it to provide a "safe" state estimate.
Key Results
- Quadrotor: The RS-EKF was much more reactive to mass changes. It achieved a 54% improvement in tracking accuracy (MSE) and a 35% reduction in total cost compared to a standard EKF-MPC setup.
- Kuka Manipulator: In 10,000 simulations with random disturbances, the RS-EKF remained closer to the reference trajectory, yielding a 32% improvement in MSE and a 22% lower mean cost.
- Solo12 Hardware: On the real robot, the RS-EKF reacted faster to external forces. When a 20N pulling force was released, the RS-EKF "underestimated" the robot's height, which caused the MPC to generate higher ground reaction forces, returning the robot to its target height much faster than the EKF. This resulted in a 50-60% improvement in average cost on hardware.
Limitations and Future Work
- Steady-State Error: Because the filter is always slightly "pessimistic," it can introduce a small steady-state error even when there is no disturbance. The robot might never perfectly reach a target because it's constantly hedging against a "worst-case" that isn't happening.
- Parameter Tuning: The risk-sensitivity parameter ($\mu$) must be tuned carefully. If $\mu$ is too large, the underlying math problem becomes "ill-posed" (the adversary becomes too powerful), and the filter can fail.
- Future Directions: The authors suggest applying this to more dynamic tasks (like running or jumping) and exploring more complex measurement noise models beyond simple Gaussians.
Summary updated