Machine learning driven self-discovery of the robot body morphology

Conventionally, the kinematic structure of a robot is assumed to be known and data from external measuring devices are used mainly for calibration. We take an agent-centric perspective to explore whether a robot could learn its body structure by relying on scarce knowledge and depending only on unorganized proprioceptive signals. To achieve this, we analyze a mutual-information-based representation of the relationships between the proprioceptive signals, which we call proprioceptive information graphs (pi-graph), and use it to look for connections that reflect the underlying mechanical topology of the robot. We then use the inferred topology to guide the search for the morphology of the robot; i.e. the location and orientation of its joints. Results from different robots show that the correct topology and morphology can be effectively inferred from their pi-graph, regardless of the number of links and body configuration.

Methods
The datasets contain the proprioceptive signals for a robot arm, a hexapod, and a humanoid, including joint position, velocity, torque, body angular and linear velocities, and body angular and linear accelerations. The robot manipulator experiment used simulated robot joint trajectories to generate the proprioceptive signals. These signals were computed using the robot's Denavit-Hartenberg parameters and the Newton-Euler method with artificially added noise. In the physical experiment, joint trajectories were optimized for joint velocity signal entropy, and measurements were obtained directly from encoders, torque sensors, and inertial measurement units (IMU). In the hexapod and humanoid robot experiments, sensor data was collected from a physics simulator (Gazebo 11) using virtual IMU sensors. Filters were applied to handle measurement noise, including low-pass filters for offline estimation and moving average filters for online estimation, emphasizing noise reduction for angular velocity measurements.