We presents our experiment videos in this website. Note that the speed of all videos are not modified the experiments are recorded as is.
I. Dynamic Box Environment
Video 1: RMP is easily stucked to difficult local minia of both sides of the box.
Video 2: HiPBI(h=50, num_rollout=100). HiPBI solves this environment in most cases. However, the long computational time is detrimental, where its previous solution is usually no longer valid. Reducing rollout horizon is an option for faster computation but it would reduce HiPBI look-ahead performance for avoiding local minima.
Video 3: HiPBOT (h=10). With efficient short rollout evaluations using the SDF collision cost (including the boundaries), together with fast Sinkhorn optimization, HiPBOT predicts good situational actions in time.
II. Extreme Dynamic Maze Environment (Stress-Test)
In this section, we show our method and the baselines' demontrations in extreme dynamic and dense environment, where the velocity/acceleration level of obstacles is capped at 100 pixels. We additionally adds some Brownian noise into the obstacle accelerations, thus changing them over time. Initially, both obstacle velocities and accelerations are uniformly randomized from 0 - 100. Note that when in collision, the agent particle changes to red.
Video 4: RMP is not able to reactive in time with extreme obstacles' velocity/acceleration.
Video 5: (Left) HiPBI (h=25, num_rollout=50). (Right) HiPBI (h=50, num_rollout=100). As expected, the poor planning rate hurts HiPBI even more in this case. A shorter horizon does not help despite having better planning rate, since it would mean less obstacle anticipation for HiPBI.
Video 6: (Left) HiPBOT (h=5). (Right) HiPBOT (h=10). HiPBOT solves this environment in most cases with very short rollout horizons. However, too short rollout horizon is also detrimental (left).
Video 7: HiPBOT (h=10). The most extreme case with 50 fast obstacles (top right corner of Fig. 4 in the paper). Although HiPBOT does not manage to solve this without collision, but it is able to reduce collisions and reach the goal.
III. Panda Environment
a) Static Case
Video 7: In most cases, RMP usually stucks since there are many local minima in high-dimension robots.
Video 8: (Left) HiPBI (h=25, num_rollout=100). (Right) HiPBI (h=50, num_rollout=100). HiPBI is able to avoid obstacles in most cases. However, we observe the oscillation behavior near the goal in the green box. We speculate that due to the normalizing constraint, HiPBI overestimates the priorities of avoidance expert policies.
Video 9: (Left) HiPBOT (h=10) success case. (Right) HiPBI (h=10) failure case. HiPBOT solves this environment with very smooth movements and more safe behaviors. However, with lacking of exploration, HiPBOT still suffers from local minima.
b) Dynamic Case
Video 9: RMP is also often stuck in dynamic obstacle cases.
Video 10: (Left) HiPBI (h=25, num_rollout=100). (Right) HiPBI (h=50, num_rollout=100). HiPBI still suffers from overestimation problem from the normalizing constraint of policy temperatures.
Video 10: (Left) HiPBOT (h=5) (Right) HiPBOT (h=10). HiPBOT also solves the dynamic environments in most cases. However, with a bit longer rollout horizons, it finds smoother path towards the goal.