How Tetris‑Inspired Blocks Enabled a 25 mm/s Micro‑Robot via Closed‑Loop AI Optimization

Problem Statement

Designing vibration‑driven micro‑robots requires morphologies that are simultaneously manufacturable, stable, and capable of converting vertical vibration into forward locomotion. Traditional pixel‑level binary encoding yields an extremely sparse set of feasible designs in the high‑dimensional discrete design space.

Block‑Based Shape Encoding

To increase the density of viable candidates, the authors replace individual pixels with the seven Tetris tetrominoes (I, J, L, O, S, T, Z). A design is represented on a 15 × 15 grid by a four‑tuple (shape, rotation, x, y). During generation the following constraints are enforced:

All placed blocks must be edge‑connected, guaranteeing a single continuous body.

The outer contour must lie within the 15 × 15 boundary.

Static‑stability criteria (center‑of‑mass within support polygon) are evaluated and only configurations that satisfy them are kept.

This “block‑organ” representation ensures that every sampled genotype can be directly fabricated by 3D printing.

Hardware‑in‑the‑Loop Optimization Loop

The closed‑loop pipeline consists of:

Algorithmic generation of a batch of candidate morphologies.

Selective laser sintering (SLS) or stereolithography (SLA) 3D printing of each candidate.

Mounting the printed part on a vibration test platform driven by a function generator → power amplifier → electrodynamic shaker.

Recording vertical displacement with a high‑speed camera (≥ 2000 fps) and extracting the average forward speed (mm / s).

Feeding the measured speed back to the surrogate model as the objective value.

The loop is illustrated in

.

Surrogate Model and Exploration Strategy

A random‑forest regressor is trained on the accumulated (genotype, speed) pairs to predict performance of unseen candidates. Exploration proceeds in two stages:

Monte‑Carlo sampling: Randomly generate a large set of feasible block configurations, evaluate them with the surrogate, and select the top‑k with high predicted speed and high prediction uncertainty (using variance from the forest).

Genetic refinement: Apply a steady‑state genetic algorithm to the selected set. Mutation consists of replacing a single block with another tetromino, rotating it, or shifting its position while preserving feasibility. Crossover is omitted to avoid breaking connectivity. Diversity is maintained by limiting similarity between individuals (e.g., Hamming distance on the block tuple list).

Experimental Results – Micro‑Robot Evolution

Thirty optimization iterations were performed. Speed measurements (average forward speed under a 30 Hz vibration) showed rapid improvement:

Iteration 2 surpassed 20 mm / s.

Iteration 26 reached a peak of 25.27 mm / s.

The speed distribution continuously shifted rightward, indicating consistent discovery of better morphologies.

The final high‑performance morphology exhibits a three‑segment architecture:

Fore‑limb: A continuous material strip at the leading edge that provides stable ground contact.

Torso: A rigid central channel that efficiently transmits vertical vibration energy toward the rear.

Tail: A partially hollow rear section that reduces drag; during vibration the tail lifts ≈ 2.1 mm, moving the contact point forward and generating net propulsion.

High‑speed imaging confirms the lever‑effect of the tail lift.

Benchmark Comparison

For the same evaluation budget, an offline random‑search baseline never exceeded 13.39 mm / s, while the closed‑loop method broke the 20 mm / s barrier in the second iteration and ultimately achieved 25.27 mm / s. This demonstrates the advantage of real‑world feedback combined with informed sampling in a highly constrained discrete space.

Generalization to Classical Combinatorial Problems

The same representation‑constraint‑closed‑loop framework was applied to three benchmark combinatorial optimization problems: Trap‑64, 0‑1 Knapsack‑250, and Max‑Cut‑128. Within a fixed computational budget the method consistently outperformed pure random search, indicating that the approach is applicable to a broad class of discrete topology‑optimization tasks such as soft actuators and mechanical metamaterials.

Reproducibility

All source code, design data, and experimental logs are deposited on Zenodo (DOI: 10.5281/zenodo.14978583). The peer‑reviewed article is available in Materials & Design, 257 (2025) 114533, DOI: 10.1016/j.matdes.2025.114533.

Key Insight

Replacing pixel‑level encoding with modular block primitives, embedding geometric and stability constraints during genotype creation, and closing the loop with physical performance measurements enable automatic discovery of manufacturable, high‑speed micro‑robot morphologies. The pipeline provides a general strategy for AI‑driven design in discrete, highly constrained engineering domains.