Self-Organized Construction by Minimal Surprise

TL;DR

The paper investigates how a minimal-surprise intrinsic drive can enable self-organized collective construction in robot swarms without explicit task definitions. It couples a pair of neural networks per robot—a forward-action network and a sensor-prediction network—evolved via a genetic algorithm, with fitness based on prediction accuracy and subsequent block pushing to create predictable environments. Results show robust emergence of structures (dispersion, pairs, lines, clusters) whose occurrence depends on robot-to-block density and grid size, and demonstrate that predefined sensor predictions can steer the swarm toward specific formations. The approach suggests a scalable, hardware-feasible path to programmable self-organization in swarm robotics, with seeds and prediction engineering offering extra control levers for future work.

Abstract

For the robots to achieve a desired behavior, we can program them directly, train them, or give them an innate driver that makes the robots themselves desire the targeted behavior. With the minimal surprise approach, we implant in our robots the desire to make their world predictable. Here, we apply minimal surprise to collective construction. Simulated robots push blocks in a 2D torus grid world. In two variants of our experiment we either allow for emergent behaviors or predefine the expected environment of the robots. In either way, we evolve robot behaviors that move blocks to structure their environment and make it more predictable. The resulting controllers can be applied in collective construction by robots.

Figures (7)

• Figure 1: Sensor model. The gray circle represents the robot. The arrow indicates its heading.
• Figure 2: Action network and prediction network. $A(t-1)$ is the robot's last action value and $A(t)$ is its next action. $T(t)$ is its turning direction. $s_0(t),\dots,s_{11}(t)$ are the robot's 12 sensor values at time step t, ${p_0(t+1),\dots,p_{11}(t+1)}$ are its sensor predictions for time step $t+1$.
• Figure 3: Best fitness of 20 independent evolutionary runs on the $20\times 20$ grid with 50 robots and 50 blocks. Medians are indicated by the red bars.
• Figure 4: Structures at the start (left) and end (rights) of a run. Robots are represented by triangles, blocks by circles. Triangles give the robots' headings.
• Figure 5: Best fitness of 20 independent runs on the $20\times 20$ grid with 20 robots and 50 blocks and predefined pairs. Medians are indicated by the red bars.