There are plenty of advanced quadruped robots that can take on dangerous tasks in remote and hazardous environments. Locomotion in rough environments remains a challenge as sensors degrade in accuracy depending on lighting, fog, water, and reflective surfaces. Relying solely on proprioception can limit locomotion speed. Researchers at Robotic Systems Lab have come up with perceptive locomotion controller for quadrupedal robots that enables them to adapt to challenging terrains for faster locomotion.
As the researchers explain, they
leverage an attention-based recurrent encoder that integrates proprioceptive and exteroceptive input. The encoder is trained end-to-end and learns to seamlessly combine the different perception modalities without resorting to heuristics. The result is a legged locomotion controller with high robustness and speed.