Unbiased Active Inference for Classical Control

Mohamed Baioumy, Corrado Pezzato, Riccardo M. G. Ferrari, Nick Hawes

Introduction

Active Inference (AIF) is a theoretical framework originating in computational neuroscience that describes biological agents as entities minimizing "variational free energy"—a proxy for surprise. In robotics, this framework has been implemented as the Active Inference Controller (AIC). While AIC offers a biologically inspired approach to control, it suffers from two significant engineering limitations: biased state estimation and implicit control modeling. These flaws lead to poor tracking performance, safety risks, and "integrator windup" in complex robotic systems.

This paper introduces the unbiased-AIC (u-AIC), a redesigned probabilistic model that decouples perception from goal-seeking behavior. By treating control as an explicit variable and removing goal-biases from state estimation, u-AIC achieves robust, stable, and accurate control suitable for real-world manipulators.

The Limitations of Standard AIC

In the standard AIC implementation, the robot's "belief" about its current state is intentionally biased toward its desired goal. This "wishful thinking" mechanism is what drives the robot to move; however, it creates a fundamental conflict. The robot never accurately perceives its actual position unless it has already reached the target. This bias results in "false positive" fault detections and prevents the system from correctly identifying external perturbations.

Furthermore, standard AIC derives control actions implicitly via the chain rule rather than modeling them explicitly. This causes the controller to behave like a simple integrator. If the robot's path is obstructed, the control effort continues to build up (windup), leading to dangerous high-velocity jerks or massive overshoots once the obstacle is removed.

Methods: The u-AIC Framework

The proposed u-AIC addresses these issues by restructuring the underlying probabilistic graphical model. The core innovation lies in moving the goal encoding from the state prior $p(x)$ to a conditional control distribution $p(u|x)$.

The u-AIC (left) moves the goal encoding from the state prior to the control distribution, unlike the standard AIC (right).

This architectural shift allows the state estimation process to remain objective and unbiased. By treating the control action $u$ as an explicit random variable, the authors can perform direct gradient descent on the free energy with respect to $u$. This transforms the AIC from a coupled, non-standard flow into a modular architecture where accurate state estimation directly informs a robust controller. This modularity also allows for the inclusion of classical control features such as feed-forward signals and control costs.

Experimental Results

The researchers validated the u-AIC using both 2-DOF simulations and real-world experiments with a 7-DOF Franka Emika Panda manipulator.

Robustness to Collisions

A critical test involved a collision scenario where the robot arm was physically blocked while moving toward a target. In the standard AIC, the belief about the state deviated significantly from reality during the block, and the control effort ($u$) ramped up continuously. Upon release, the AIC exhibited a massive overshoot of approximately 43.9%.

Simulation results comparing AIC and u-AIC during a collision (shaded region). u-AIC maintains accurate state estimation and avoids the control windup seen in standard AIC.

In contrast, the u-AIC maintained an accurate, unbiased estimate of its position throughout the collision. Because it "knew" it was stuck, it did not suffer from integrator windup, resulting in a much safer and more stable recovery with only ~11.6% overshoot.

Tracking and Accuracy

In reference tracking tasks, such as following a sinusoidal path, the standard AIC failed to converge due to its inherent perceptual bias. The u-AIC, however, demonstrated superior accuracy, with state estimation Root Mean Square Error (RMSE) being two orders of magnitude lower than that of the standard AIC.

Conclusion

The u-AIC represents a significant step forward in making Active Inference viable for industrial and collaborative robotics. By removing the "wishful thinking" bias inherent in standard AIC, the authors have created a controller that is not only more accurate but also significantly safer in unpredictable environments. Future work will focus on integrating more complex dynamic models and automated parameter tuning to further enhance performance against modern benchmarks like Model Predictive Control (MPC).

Summary updated

BibTeX
@inproceedings{baioumy2022unbiased,
  author = {Mohamed Baioumy and Corrado Pezzato and Riccardo M. G. Ferrari and Nick Hawes},
  title = {{Unbiased Active Inference for Classical Control}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2022},
  doi = {10.1109/iros47612.2022.9981095},
}

Loading related content...

Figure