alexpostnikov / backprop-mpdm_torch Goto Github PK

first transition function is moving to wrong direction?

ped_cov = calc_linear_covariance(x[0])
for ped in ped_cov.keys():
for timestemp in range(ped_cov[ped]["start_step"]+1, ped_cov[ped]["end_step"]):
cov_prev = ped_cov[ped]["cov"][timestemp-1]
pose_prev = ped_cov[ped]["pose"][timestemp-1]
pose_cur = ped_cov[ped]["pose"][timestemp]

i`ve got IndexError

That way of calculation can optimize speed of calculation

File "scripts/cov_prediction/model.py", line 64, in
ped_cov = calc_linear_covariance(x[0])
File "scripts/cov_prediction/model.py", line 30, in calc_linear_covariance
peds[ped]["vel"] = peds[ped]["pose"][1] - peds[ped]["pose"][0]
IndexError: index 1 is out of bounds for axis 0 with size 1

hsfm-sigmann Covariances are too narrow sometimes

pedestrians with only 1 position in dataset ignorred at now.
see Validator.py lines 25-28
# rm peds with only 1 pose
for ped in peds:
if len(peds[ped]["pose"])<2:
del peds[ped]

seems lile the "Left" and "Right" policies have bigger priorities than "Go-solo"

make goal distrib like (rotated) ellipsoid

that enchansment will help new users

ориентация пешехода должна учитываться в модели движения. Возможно и робота тоже. Стоит сверить с cpp кодом и перечитать статьи

check probabilities during cost calculations

make gradients propagate till initial configuration for social-lstm.

I found in cpp code this idea:

switch(distance_to_nearest_ped):
high_dist: robot_state_=FREE_ZONE_NAVIGATING || PORA_FORSITb
mid_dist: robot_state_=MID_ZONE_NAVIGATING
low_dist: robot_state_=INNER_ZONE_NAVIGATING

robot_state_=IDLE - stop navigation. Set this mode manual for stop moving

robot_state_ need to controll maximum robot speed.

That options can improve expireance of debugging

need to add new parametr onto model calculation

also include there all used articles

by analogy with the calculations of the position

    self.a = 0.001
    self.b = 3

in Param.py

prediction speed of pedestrians with some delta times may become like a fur-tree.