Minimal Sensor Quadruped Locomotion Without Contact Models via Reinforcement Learning

Quadruped robots are increasingly deployed in environments where sensing is degraded, communication is restricted, and terrain properties are uncertain conditions common in planetary exploration, subterranean mapping, and disaster-response scenarios. These missions demand locomotion controllers that are robust, lightweight, and capable of operating with minimal hardware dependencies. In this work, we present a proprioceptive-only quadruped locomotion framework that eliminates the need for exteroceptive sensors, explicit contact models, or detailed terrain information. Our approach leverages the Soft Actor-Critic (SAC) reinforcement learning algorithm to learn adaptive gait behaviors solely from inertial and joint measurements. By providing only a simple reference foot-strike schedule during training, the policy implicitly learns contact patterns and ground interactions without requiring force sensors, foot switches, or model-based contact dynamics. This results in a controller with significantly reduced computational and hardware complexity, enabling deployment on platforms with strict payload, power, and reliability constraints. Experimental results in simulation demonstrate that the learned controller maintains stable locomotion across a variety of uneven and unpredictable terrains with robustness and adaptability. The proposed framework highlights the potential of minimal-sensor, model-free RL policies for enabling resilient quadruped mobility in environments where conventional sensing and planning pipelines are infeasible.