Environment-Centric Active Inference

TL;DR

The paper addresses the limitations of agent-centric Markov Blanket designs in active inference, which can fail to respond to unintended changes in the environment. It proposes environment-centric active inference (EC-AIF), defining the Markov Blanket from the environment using a grid-based observable space and a what/where decomposition to generate observations $o$, hidden states $s$, and policies $\pi$ through a FlowECAIF loop within the generative model $p(o,s,\pi)$. The key contributions include a formalization of EC-AIF, a scalable environment-driven MB construction, and validation on a dual-robot object-transport task that demonstrates robust adaptation to changes in target position and the orientation of another robot. The results show that EC-AIF enables robust open-environment inference and cooperative robotics without predefined agent boundaries, with potential for broader applications in multi-agent systems.

Abstract

To handle unintended changes in the environment by agents, we propose an environment-centric active inference EC-AIF in which the Markov Blanket of active inference is defined starting from the environment. In normal active inference, the Markov Blanket is defined starting from the agent. That is, first the agent was defined as the entity that performs the "action" such as a robot or a person, then the environment was defined as other people or objects that are directly affected by the agent's "action," and the boundary between the agent and the environment was defined as the Markov Blanket. This agent-centric definition does not allow the agent to respond to unintended changes in the environment caused by factors outside of the defined environment. In the proposed EC-AIF, there is no entity corresponding to an agent. The environment includes all observable things, including people and things conventionally considered to be the environment, as well as entities that perform "actions" such as robots and people. Accordingly, all states, including robots and people, are included in inference targets, eliminating unintended changes in the environment. The EC-AIF was applied to a robot arm and validated with an object transport task by the robot arm. The results showed that the robot arm successfully transported objects while responding to changes in the target position of the object and to changes in the orientation of another robot arm.

Figures (9)

• Figure 1: Concept of (a) conventional Markov Blanket (b) Markov Blanket of EC-AIF.
• Figure 2: Decision process of observation, hidden states, and actions variables in EC-AIF.
• Figure 3: Robots in Experiment (a) UR5e (b) COBOTTA.
• Figure 4: Scenarios of (a) the target position for object transport changes, and (b) another robot's orientation changes. Each circle represents a grid point comprising $where$. UO is the origin position of UR5e, CO is the origin position of COBOTTA, and Int. is the intermediate position between UR5e and COBOTTA. The light blue circles (P1, UO, P6, P7, P8, P11, P12, P13) are grid points in the reach range of UR5e, the blue circles (CO, P5, P10, P15) are grid points in the reach range of COBOTTA, and the purple circles (Int., P9, P14) are grid points in the reach range of both robots. In the experiment, the grid points are not evenly distributed, and the position of the points is shifted due to the constraints of the mechanism in which the robot is installed.
• Figure 5: Transition of the selected action when the target position is within the reach of UR5e by (a) normal AIF and (b) EC-AIF.