Fast Kinodynamic Planning on the Constraint Manifold
with Deep Neural Networks

Planning under complex kinodynamic constraints is a challenging task, particularly when reactive motions are needed and it is not possible to employ advanced optimization schemes due to the limited computation time allowed to take actions. While novel learning methods have shown promising solutions for these tasks, proposed approaches still lack comprehensive handling of complex constraints, such as planning on a lower-dimensional manifold of the task space. In our work, we introduce a novel learning-to-plan framework that exploits the concept of constraint manifold combined with neural planning methods. Our approach generates plans satisfying an arbitrary set of constraints while requiring minimal computation time, namely the inference time of a neural network. The short planning time allows the robot to plan and replan reactively, making our approach suitable for dynamic environments. We validate our approach on two simulated tasks and in a demanding real-world scenario, where we use a Kuka LBR Iiwa 14 robotic arm to perform the hitting movement in robotic Air Hockey.

I. Experiments in simulation

We evaluated our solution in two simulated experiments:

• moving a heavy vertically-oriented object

• high-speed hitting in an Air Hockey game

III. Trick shots

Our planner, thanks to its unparalleled properties i.e. short deterministic planning time and its ability to plan taking into account initial velocity and control, is able to smoothly replan from the actually followed trajectory to the new one. This allows us to perform some trick shots, that require rapid replanning. 

IV. The hitting experiment videos

We include here the video recording of our hitting experiments (no cuts, except for the camera battery change)