MPC-MPNet-Path
MPC-MPNetTree
Model Predictive Motion Planning Network (MPC-MPNet) is a neural network-based kinodynamic planning algorithm capable of solving Kinodynamic Motion Planning tasks in seconds with less variability and comparable path quality against a state-of-the-art algorithm that takes minutes to generate solutions with high variability. MPC-MPNet utilize CNN-based neural network to learn environment embedding and predict waypoints satisfying kinodynamic constrains. Our experiments show that MPC-MPNet generalizes to unseen tasks and adapts to high-dimensional problems such as Quadrotor with high success rate and high-quality solutions. MPC-MPNet is fully parallelizable and available for GPU computation, achieving practical completeness and optimality. Our algorithms show theoretical asymptotic optimality when SST is used as a backup planner. Moreover, MPC-MPNet provides a generalized framework for the KMP planner, where multiple variants can easily be adapted.
We evaluate MPC-MPNet to plan motion for four underactuated dynamic systems: Acrobot, Cartpole, Car and Quadrotor. The environments are random generated and the start and goal states are unseen during training.
time=2.967s cost=5.52
time=5.19s cost=5.74
time=27.82s cost=7.20
time=0.413s cost=6.98
time=8.18s cost=5.50
time=20.49s cost=6.86
time=15.53s cost=6.02
time=4.00s cost=6.31
time=25.44s cost=9.50
time=2.12s cost=3.49
time=0.80s cost=3.00
time=11.45s cost=5.10
time=6.48s cost=48.79
time=11.63s cost=52.01
time=43.83s cost=49.03
time=4.91s cost=33.71
time=7.83s cost=46.13
time=22.86s cost=70.51
time=3.75s cost=14.66
time=7.70s cost=11.45
time=363.83s cost=15.78
time=0.65s cost=7.416
time=1.71s cost=8.69
time=107.56 cost=15.044