Kevin S. Tracy
Description: Differentiable Continuous Collision Detection for Convex Sets
nerf (195) d-nerf (90) nerfies (89)
--> Colab Your browser does not support the video tag. Collision-avoidance with discrete collision detection (left) and continuous collision detection (right). These collision constraints are fully differentiable and work with of-the-shelf trajectory optimizers. Abstract Typical discrete collision detection can only consider robots and environments in their static configurations at discrete time steps. For sequences in which either the robot or the environment moves, discrete collision detection is unable t
We solve for continuous collision information in an optimization framework. Instead of solving for closest point between pairs of convex bodies, we instead parametrize the continuous collision problem with the following two variables:
In order to differentiate our collision detection problem with respect to the configurations/geometry of the two objects, we use a clever differentiable optimization technique.
Kevin S. Tracy