Blind Locomotion

A locomotion approach that uses no camera or lidar, relying only on joint and IMU signals to walk.

Blind locomotion means a legged robot walks without using external sensing such as a camera or lidar, relying only on proprioceptive signals such as joint encoders and the IMU (inertial measurement unit), inferring the terrain from foot-contact feedback alone. A landmark example is the ANYmal quadruped controller from Marco Hutter's group at ETH Zurich, published in Science Robotics in 2020: the neural network reads only a sequence of proprioceptive signals, trained with teacher-student distillation in simulation and then transferred zero-shot to real-world terrain such as mud, snow, gravel, dense vegetation, and flowing water. In 2021, Jonah Siekmann and colleagues (RSS) used a similar approach to get the biped robot Cassie to climb a real staircase using only proprioception. The advantage of blind locomotion is that it is unaffected by vision failing, in darkness, smoke, or tall grass, and it avoids the latency and error that come with mapping; the drawback is that it can only react after a foot has already landed, making it hard to plan ahead for a high step or a gap. Because of this, it is often combined with perceptive locomotion, serving as a fallback when vision is unreliable.

ExampleAn ANYmal quadruped, without looking at the ground at all and relying only on its leg joint and IMU signals, walks through mud, snow, and swift-flowing water.

Also called

Proprioceptive Locomotion

Related

Perceptive Locomotion · Proprioception · Legged Locomotion · Rough-terrain Locomotion · Teacher-Student Distillation · Sim-to-Real Transfer

Sources

Learning Quadrupedal Locomotion over Challenging Terrain (Lee et al., Science Robotics 2020)
Blind Bipedal Stair Traversal via Sim-to-Real Reinforcement Learning (Siekmann et al., RSS 2021)