Rapid Discovery of Robot Designs via Model Predictive Control

Overview

Through the process of evolution, natural organisms have developed distinct morphologies suited to their environment and mode of survival. With this as inspiration, recent work has studied generating robot designs de novo with evolutionary algorithms. To measure fitness, these works commonly use reinforcement learning (RL) to first train a controller for a given task and robot design. Although this has shown success, evolution with RL is costly, and can be computationally prohibitive when searching over a large population size or scaling the number of generations. In this work, we introduce Evolutionary Predictive Sampling (EPS): an evolutionary algorithm that uses a derivative-free variant of model predictive control (MPC) in the inner loop. Unlike RL, MPC requires no prior training on a fixed morphology, making it uniquely suited to evaluating candidate designs with minimal simulation. This enables our method to use 80x fewer simulation steps per morphology evaluation. We evaluate our method on four tasks: traversing rough terrain, walking on flat ground, reorienting a cube, and transporting an object to a goal, where the evolutionary process produces consistent gains in population fitness over 50 generations. On the primary task of rough terrain locomotion, EPS achieves 63% higher task reward compared to an RL baseline with a fixed compute budget of 50 billion simulation steps. We believe our work will help accelerate future research in robot co-design by significantly reducing the compute and wall-clock time required to compare a large number of morphologies.

Competition history

• CSEF 2026 Applied Mechanics (Senior Division) · Entry S-02-17

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Engineering projects
• CSEF Engineering projects
• Every CSEF 2026 project