Budget-Friendly Robot Dog Masters Rough Terrain

Researchers built a control system for a low-cost four-legged robot. The system blends central pattern generators with a multi-agent reinforcement learning layer. This combination lets the robot walk over obstacles, ramps, stairs, and uneven ground.

Each leg acts as an individual agent that shares a single policy network. A small residual module fine-tunes the basic gait pattern. The entire network contains roughly thirty-nine thousand parameters and fits in about one hundred fifty-five kilobytes.

The robot does not require a central critic, external cameras, or torque sensors on board. Those components are handled in simulation or offline. This simplifies the hardware and keeps the cost low.

Tests showed the robot traversing obstacles, slopes, stairs, and rough terrain. The system proved more stable, used less energy, and adapted better than earlier controllers. These gains appeared both in virtual trials and on a physical platform.