A quadrupedal adaptive foot has been developed, and it has shown the ability to adapt to rough terrain shapes and improve grip. This foot, namely SoftFoot-Q, is also equipped with IMUs to enable pose estimation and haptic exploration.

In the context of robotic environmental monitoring of natural habitats, quadrupedal robots are the perfect candidates to tackle the challenging terrains involved. To further improve balancing performance, adaptive feet can be employed to maximize the grip for the interaction with the environment. Dynamic locomotion gaits are preferable to static gaits to achieve higher velocities during navigation. Moreover, highly dynamic motions are essential to succeed in these scenarios. Data-driven methods can be used to synthesize locomotion controllers capable of addressing such complex problems. In particular, neural-network-based policies can be trained offline in simulation, via reinforcement learning, and can then be deployed on the real robot. Domain randomization is crucial to bridge the sim2real gap and highly parallel GPU-accelerated physics simulators are helpful to speed up the learning process. With this setup, it is possible to train high performance end-to-end policies, which take raw data from on-board sensors as input and output joint position commands. These commands are then sent to motor-level PD controllers, which are in charge of tracking the motion. The addition of haptic terrain perception, by means of the IMU sensors mounted on the adaptive feet, helps to further improve locomotion and balancing performance.

Mastering rough terrain is challenging for robots due to its unpredictability. Traditional methods rely on precise foot placements, but recent advancements in quadruped robot feet offer diverse shapes and high grip. This work introduces a foothold optimization method using polynomial approximation and a CNN trained on simulated data to predict the cost for each candidate foothold. The method efficiently searches for optimal contact points for various foot shapes.

The validation tests were performed in four different scenarios: forest, grassland, dune and screes.