Adaptive motion planning for legged robots in unstructured terrain using deep reinforcement learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Legged robots offer superior mobility over unstructured terrain compared to wheeled platforms.
• Motion planning for legged robots is complex due to unstable ground, limited sensing, and dynamic body movements.
• Classical control methods struggle with unpredictable changes in terrain and friction.

Purpose of the Study:

• To develop an adaptive locomotion framework for quadrupedal robots capable of navigating diverse and unpredictable environments.
• To train a controller using deep reinforcement learning (DRL) that can adapt to varying terrain conditions and dynamics.
• To ensure the learned locomotion policy is robust and generalizable across different simulation environments.

Main Methods:

• A deep reinforcement learning (DRL) controller was trained using the proximal policy optimization (PPO) method entirely in simulation (Webots).
• Training employed a curriculum mechanism, starting with flat terrain and progressively introducing slopes, rough surfaces, and friction variations.
• The controller learned to select footholds, regulate body attitude, and minimize slip based on joint states, body attitude, and local terrain height, without explicit terrain labels.

Main Results:

• The trained controller demonstrated emergent behaviors such as wider stances on slopes and increased step height on rocks for improved stability.
• The framework achieved high success rates (94.6% in Webots, 91.2% in PyBullet) without retraining and recovered from external disturbances within 1.6 seconds.
• The system learned stable, energy-efficient locomotion with minimal foot slippage across a range of friction coefficients.

Conclusions:

• The proposed adaptive locomotion framework effectively enables quadrupedal robots to navigate challenging, unstructured terrains.
• Deep reinforcement learning provides a viable method for training robust and adaptable locomotion policies for legged robots.
• The learned behaviors enhance robot stability, energy efficiency, and resilience, paving the way for applications in search-and-rescue and exploration.