The availability of inexpensive 3D-printed quadrupedal robots motivates the development of learning-based methods compatible with low-cost embedded processors and position-controlled hobby servos. In this work, we show that a linear policy is sufficient to modulate an open-loop trajectory generator, enabling a quadruped to walk over rough, unknown terrain, with limited sensing. The policy is trained in simulation using randomized terrain and dynamics and directly deployed on the robot. We show that the resulting controller can be implemented on resource-constrained systems. We demonstrate the results by deploying the policy on the OpenQuadruped, an open-source 3D-printed robot equipped with hobby servos and an embedded microprocessor.

We introduce a distributed finite element algorithm that allows swarms of mobile robots to persistently monitor environmental quantities such as temperature or salinity. The robots deploy themselves into the environment, covering the domain and dividing it into nonoverlapping regions. Each robot estimates the environment over its own region using local measurements and communication with nearby robots. The algorithm ensures that each robot's estimate constitutes a piece of a global estimate that spans the entire domain, fuses the whole swarm's measurements, and accounts for the spatial correlation between measurement and estimation locations. By incorporating spatial correlation without requiring the transmission of measurements or measurement locations, the algorithm decouples its communication requirements from the spatial statistics of the environment and enables robots with fixed capabilities to monitor environments with different spatial correlation lengths. Analysis and simulation demonstrate that, as the number of robots increases, the memory and communication requirements of each individual robot decrease until reaching a minimum, after which the resolution of the environmental model increases. Additional robots, therefore, add computational resources to the swarm rather than introducing extra computational burdens.

Average consensus estimators enable robots in a communication network to calculate the mean of their local inputs in a distributed manner. Many distributed control methods for robot swarms rely on these estimators. The performance of such estimators depends on their design and the network topology. For mobile sensor networks, this topology may be unknown, making it difficult to design average consensus estimators for optimal performance. We introduce a design method for proportional-integral (PI) average consensus estimators that decouples estimator synthesis from network topology. This method also applies to the more general internal model (IM) estimator, yielding extended PI estimators that improve convergence rates without increasing communication costs. In simulations over many geometric random graphs, the extended PI estimator consistently reduces the estimation error settling time by a factor of five.