Mohammad Salah Uddin
Legged robots are ideal for navigating unstructured terrain. Unlike wheeled robot platforms, legged robots maintain mobility over rocks, slopes, and uneven surfaces. However, motion planning in such environments remains challenging. The robot must handle unstable ground, limited sensing, and complex body dynamics. Classical control methods often fail to adapt when terrain and friction change unexpectedly. This paper presents an adaptive locomotion framework for quadrupedal robots. The controller is trained entirely in simulation using deep reinforcement learning (DRL). The DRL is trained using proximal policy optimization (PPO) method. During training, the robot observes joint states, body attitude, and local terrain height information. Training follows a curriculum mechanism. The robot starts on flat terrain. The difficulty gradually increases with slopes, height noise, rough surfaces, and friction variation. The reward function encourages forward progress, stability, low foot slip, and energy-efficient motion. Training occurs in Webots with wide domain randomization ranges. Policies learn to select footholds, regulate body attitude, and minimize slip without explicit terrain labels or gait scripts. Several behaviors emerge through learning. These include wider stance on slopes, increased step height on rocks, and hip abduction for balance recovery. To ensure robustness, we apply a multi-step validation process. These include cross-simulator evaluation, domain randomization Monte Carlo tests, physics consistency tests and adversarial disturbances. The controller achieves 94.6% success in Webots and 91.2% in PyBullet without retraining. It recovers from external pushes within 1.6 seconds. Results confirm that the proposed method learns stable, energy-efficient walking with minimal foot slippage across friction coefficients [Formula: see text]. This capability supports deployment in search-and-rescue, off-road mobility, and exploration tasks in unknown terrain.