General improvements

• Add in some hand-crafted trajectories into the mix via https://arxiv.org/pdf/2401.09241 (i.e. drive to goal, stop, slow, fast, even other algorithms perhaps, etc with samples)
• Allowing the std of the velocities to be weighted by the speed of the robot (and/or speed of the control sample) or by the prediction horizon (less noise further out; Ornstein-Uhlenbeck processes)

Improved smoothness options

• Test using a Log-Normal sampling distribution as established in Log-MPPI paper. Handles less noise for similar exploration of space, reducing sampling noise needing to process.
• Add in the derivative SMPPI term (mutually exclusive with 1% sampling). Attempts to smooth using differential changes in control
• Use the described Tsallis VI variation described below to weight only the best samples in averaging
• Mostly fine, just needs minorly better path post processing - increase the SG window size or order, recursively call it, or consider other options
• Sample noise within the feasible acceleration space, not the velocity space to make sure all trajectories are feasible
• Adding a low-pass filter on each iteration update
• Add smoothness penalty function the penalizes large changes between iterations

Improvements in speed potentially

• GPU for trajectory rollouts or critic analysis
• Precomputation and otherwise optimization in a specialized collision checker object

(supporting std change over prediction horizon)