Autonomous Modification of Unstructured Environments: Modeling System Parameters (NERC'19)

Given a terrain height function (Fig. 1), the Minimal Additive Ramp Structure (MARS) provides the upper bounds for the least additive construction for building perfect motion support structures and the MARS Gap provides the navigable substructure to build. Physics simulators aim to reduce the simulation-to-reality gap by employing accurate shape description of the objects. We utilize a sample-based physics simulation method that models the uncertainties in the world description, and performs approximate forward simulations using samples from the models. Given a stone shape distribution and terrain height function, multiple stone meshes are evaluated in order to find optimal deposition poses.

In order to make this project work as designed, we need to perform a large amount of simulation runs. Since the terrain and stones are both noisy representations of the real world, we do not want to perform accurate, computationally expensive simulations. Hence, we look at the tradeoff between accuracy and speed for different simulation parameters, to tune the simulator given the terrain height function and the object shape distribution. We use pybullet for rigid body physics simulations. The controllable parameters are given by the tuple (c_sim_time, c_iters, c_max_faces). We perform parameter tuning by comparing the performance and run times of the best possible setting, (0.004, 2000, 1500), with a mean run time of 58 seconds. Our parameter search favors speed over accuracy; it chooses a value that does not "deviate" greatly from the best setting, and has considerably lower run times. The parameter tuple (0.09, 300, 250) is selected, with an average run time of 9 seconds.

* This work has accepted to NERC'19, supported by NSF#184634 and CAPES #013584/2013-08.

** This project was developed while affiliated to the Department of Computer Science & Engineering, at University at Buffalo, The State University of New York.