Numerical Instabilities in Python Bindings #240
Description
I came across weird/inconsistent behavior when using ruckig for trajectory generation in Python. I tracked down the error and created a minimum example that reproduces the error:
import pytest
from ruckig import InputParameter, Result, Ruckig, Trajectory
start_1 = [0.6900046551907393, 1.867349375974835e-06, 0.6670294289887627, 0]
start_2 = [0.6900046551918803, -1.1288094259602022e-16, 0.6670294289873331, 0.0]
bad_target = [0.6900046551907393, 1.867349375974835e-06, 0.6670294289887627, 0.9999999999999998]
good_target = [0.6900046551918803, -1.1288094259602022e-16, 0.6670294289873331, 1.0000000000000002]
@pytest.fixture
def input_params():
inp = InputParameter(4)
inp.max_velocity = [0.25, 0.25, 0.25, 1]
inp.max_acceleration = [0.5, 0.5, 0.5, 0.5]
inp.max_jerk = [10, 10, 10, 0.5]
inp.current_velocity = [0, 0, 0, 0]
inp.current_acceleration = [0, 0, 0, 0]
inp.target_velocity = [0, 0, 0, 0]
inp.target_acceleration = [0, 0, 0, 0]
return inp
@pytest.fixture
def otg():
return Ruckig(4)
@pytest.mark.parametrize("start", [start_1, start_2])
@pytest.mark.parametrize("target", [good_target, bad_target])
def test_ruckig(start, target, input_params, otg):
input_params.current_position = start
input_params.target_position = target
trajectory = Trajectory(4)
result = otg.calculate(input_params, trajectory)
assert result == Result.WorkingThe above test succeeds for both runs with the good_target, but fails for both instantiations of the bad_target with a RuckigError:
E ruckig.RuckigError:
E [ruckig] error in step 1, dof: 3 input:
E inp.current_position = [0.6900046551907393, 1.867349375974835e-06, 0.6670294289887627, 0]
E inp.current_velocity = [0, 0, 0, 0]
E inp.current_acceleration = [0, 0, 0, 0]
E inp.target_position = [0.6900046551907393, 1.867349375974835e-06, 0.6670294289887627, 0.9999999999999998]
E inp.target_velocity = [0, 0, 0, 0]
E inp.target_acceleration = [0, 0, 0, 0]
E inp.max_velocity = [0.25, 0.25, 0.25, 1]
E inp.max_acceleration = [0.5, 0.5, 0.5, 0.5]
E inp.max_jerk = [10, 10, 10, 0.5]
The error message is a bit confusing, because it says dof: 3 , whereas everything is clearly instantiated for 4 dof. I don't see why the behavior between these similar targets should be so different. If it is not a bug, a more helpful error message would be nice. I ran these tests on version 0.15.3 of the ruckig python bindings, but also going back to an older version (i.e. before the switch to nanobind) did not change the error.