Why the domain fits
Classical robotics separates state estimation (Kalman/particle filters), control (LQR, MPC, PID), and, increasingly, reinforcement learning, each solved as its own module. Active inference instead casts all three as approximate Bayesian inference under a single generative model, minimizing variational free energy for perception and learning and expected free energy — which trades off epistemic (information-gain) and pragmatic (goal-satisfaction) value — for action selection. Lanillos and colleagues argue this conceptual unification is exactly what makes robots ideal testbeds: they must maintain beliefs about hidden variables (body configuration, object states, dynamics) from ambiguous sensor streams and act to keep observations within their model's expected range. Millidge et al.'s analysis of the control-as-inference literature further shows that under certain assumptions, active inference policies reduce to standard optimal control or RL solutions, clarifying how existing controllers can be read as special cases of free energy minimization.
Application patterns in the literature
Generative models in robotics implementations are typically structured around known physics and kinematics rather than learned end-to-end: humanoid body-perception work encodes forward kinematics and multimodal (visual/proprioceptive) observation models, while manipulator controllers encode joint dynamics, inertia, and contact forces. Expected free energy then drives policy selection — Pezzato, Ferrari, and Hernández's adaptive manipulator controller (published in IEEE Robotics and Automation Letters) adapts online to dynamic uncertainty and sensor faults, and Baioumy et al.'s ball-balancing agent shows expected free energy inducing genuinely exploratory, curiosity-like behavior. Precision weighting — adjusting trust in sensory channels or priors based on inferred reliability — recurs across this work, notably in Ferrari's group's beta-residuals scheme for detecting and compensating faulty sensors. Hierarchical multi-level architectures are discussed conceptually (e.g., task-space preferences over joint-space actions) but concrete hierarchical implementations remain rare and largely confined to simulation.
Key projects and tools
pymdp is the report's one named general-purpose software library, an open-source Python package for discrete-state (POMDP) active inference introduced via a PubMed-indexed step-by-step tutorial. Riccardo Ferrari's group at Delft University of Technology has produced the most sustained manipulator-control body of work — the adaptive controller, integrated state-estimation/control/learning (ICRA 2021), unbiased active inference for classical control (IROS 2022), and beta-residual fault tolerance — with code referenced on his research site. Pablo Lanillos and collaborators (TU Munich and Japan-based partners) lead the humanoid line: body perception and reaching on a physical humanoid, and mirror self/other distinction combining active inference with neural network forward models. Bristol Robotics Lab and the Active Touch Laboratory contribute Bayesian active-sensing tactile exploration systems (fingertip arrays, whiskers) that embody active inference principles without always formalizing them as such, and Verses AI's blog post signals early industrial interest, though it is explicitly not peer-reviewed.
Open problems
The report identifies five unresolved areas needing further evidence: scalability of expected-free-energy evaluation to high-dimensional, real-time robots (addressed only in simulation so far by the "Scaling active inference" preprint); principled generative-model specification that balances hand-crafted physics with learned components; systematic head-to-head benchmarking against classical controllers and deep RL, for which no standardized benchmark suite (analogous to MuJoCo-style RL suites) yet exists; safety and robustness under distributional shift, where Zorzi et al.'s DR-FREE framework (robustifying free energy minimization over an ambiguity set of models) is proposed but not yet applied to physical robots; and institutional adoption, since active inference toolchains remain far less mature than ROS or deep-learning stacks. The report explicitly cautions that many claims rest on qualitative or proof-of-principle results rather than systematic comparison.
Reference Backbone
Christopher L. Buckley, Chang Sub Kim, Simon McGregor, Anil K. Seth (2017). The free energy principle for action and perception: A mathematical review. Journal of Mathematical Psychology. DOI: 10.1016/j.jmp.2017.09.004. Thomas Parr, Giovanni Pezzulo, Karl J. Friston (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press. Giovanni Pezzulo, Francesco Rigoli, Karl J. Friston (2017). Active Inference, homeostatic regulation and adaptive behavioural control. Progress in Neurobiology. DOI: 10.1016/j.pneurobio.2017.08.001. Pablo Lanillos, Cristian Meo, Corrado Pezzato, et al. (2021). Active Inference in Robotics and Artificial Agents: Survey and Challenges. arXiv. DOI: 10.48550/arXiv.2112.01871. Lancelot Da Costa, Thomas Parr, Noor Sajid, Sebastijan Veselic, Victorita Neacsu, Karl J. Friston (2020). Active inference on discrete state-spaces: A synthesis. Journal of Mathematical Psychology. DOI: 10.1016/j.jmp.2020.102447.