kth-sml / ltl_automaton_core Goto Github PK

The dynamic parameter "replan_on_unplanned_move" exists but does not seem implemented yet in the planner node.

Because TS state is a latched topic, if initial state is given from agent, the TS state callback will also get the same state.
This could trigger re-planning or other issues.

Could be solve by having the callback ignore the first TS state if the parameter for initial state from agent in on.

The code includes compiled binaries of the LTL2BA software. However this was compiled on 64bits desktop computer (x86_64) and doesn't work on some onboard computer with different CPU architecture (ARM, ...).

LTL2BA should therefor be compiled at the same time as the ROS package, maybe using a ROS wrapper package.

Self-loop on TS nodes seems to be critical to trap detection (see issue here).

For now, those self-loop are defined at the TS config file level. Should they always be included in the TS build process? That would only make sense to leave them in the config files if there are specific configurations where we want to avoid it.

There seems to be a bug where re-planning is triggered at each new suffix execution. Needs to be checked.

For exampe, if a parallel dimension of states is defined as:

package_state:
        ts_type: "package_state"
        initial: "empty"
        nodes:
            empty:
                connected_to:
            arrived:
                connected_to:

then the following error is thrown:

processing model dimension package_state
Traceback (most recent call last):
  File "/home/tusike/catkin_ws/src/ltl_automaton_core/ltl_automaton_planner/nodes/planner_node.py", line 385, in <module>
    ltl_planner_node = MainPlanner()
  File "/home/tusike/catkin_ws/src/ltl_automaton_core/ltl_automaton_planner/nodes/planner_node.py", line 59, in __init__
    self.build_automaton()
  File "/home/tusike/catkin_ws/src/ltl_automaton_core/ltl_automaton_planner/nodes/planner_node.py", line 166, in build_automaton
    state_models = state_models_from_ts(ts_dict, self.initial_state_ts_dict)
  File "/home/tusike/catkin_ws/src/ltl_automaton_core/ltl_automaton_planner/src/ltl_automaton_planner/ltl_automaton_utilities.py", line 49, in state_models_from_ts
    for connected_node in state_model_dict['nodes'][node]['connected_to']:
TypeError: 'NoneType' object is not iterable

If using more than 2 TS in the agent model, the guard doesn't work properly.

The problem is that graph product is done iteratively by doing the product of 2 TS at a time, including node generation and building transitions. The guard doesn't work because it is not checked always against the final node label.

One way to solve the issue would be to build all nodes of the product at once, and then only build the transitions.

[INFO] [1615043700.323179]: LTL automaton Turtlebot node: waiting for turtlebot move base action server...
[INFO] [1615043700.626788]: LTL Planner: --- Planning in progress (static) ---
[INFO] [1615043700.628346]: LTL Planner: Hard task is: ([] (uncharged -> <> charged)) && ([] <> loaded) && [](loaded -> <> unloaded) && [](!r23 && !r29)
[INFO] [1615043700.630186]: LTL Planner: Soft task is:
[INFO] [1615043700.699133]: LTL Planner: full Buchi constructed with 10 states and 64 transitions
[INFO] [1615043700.984835]: LTL Planner: full model constructed with 76 states and 266 transitions
[INFO] [1615043700.986108]: LTL Planner: initial state in TS is set([('r36', 'unloaded', 'charged')])
[INFO] [1615043701.707568]: LTL Planner: full product constructed with 760 states and 5810 transitions
[INFO] [1615043703.264106]: LTL Planner: Prefix plan: []
[INFO] [1615043703.265290]: LTL Planner: Suffix plan: ['goto_r35', 'goto_r34', 'pick', 'goto_r35', 'goto_r32', 'goto_r33', 'goto_r30', 'goto_r27', 'goto_r24', 'goto_r21', 'drop', 'goto_r24', 'goto_r27', 'goto_r30', 'goto_r33', 'goto_r36']
[INFO] [1615043703.266800]: LTL Planner: --- Planning successful! ---
Traceback (most recent call last):
File "/home/sml/catkin_ws/src/ltl_automaton_core/ltl_automaton_planner/nodes/planner_node.py", line 390, in
ltl_planner_node = MainPlanner()
File "/home/sml/catkin_ws/src/ltl_automaton_core/ltl_automaton_planner/nodes/planner_node.py", line 44, in init
self.build_automaton()
File "/home/sml/catkin_ws/src/ltl_automaton_core/ltl_automaton_planner/nodes/planner_node.py", line 157, in build_automaton
self.ltl_planner.optimal()
File "/home/sml/catkin_ws/src/ltl_automaton_core/ltl_automaton_planner/src/ltl_automaton_planner/ltl_tools/ltl_planner.py", line 80, in optimal
self.next_move = self.run.pre_plan[self.index]
IndexError: list index out of range
[ltl_planner-1] process has died [pid 2064, exit code 1, cmd /home/sml/catkin_ws/src/ltl_automaton_core/ltl_automaton_planner/nodes/planner_node.py __name:=ltl_planner __log:=/home/sml/.ros/log/e22ec8f0-7e83-11eb-badc-f48e38896d95/ltl_planner-1.log].
log file: /home/sml/.ros/log/e22ec8f0-7e83-11eb-badc-f48e38896d95/ltl_planner-1*.log

For some specific applications, receiving twice the same TS state could make sense. Right now the hard-coded behavior is to ignore and print a warning when the received TS state is the same as the previous one. It would be interesting to have a parameter to accept or not repeating TS states

In some cases it could be useful to limit actions from one TS to only a few states of another TS. For example a pick/drop action could only be carried out in some 2D regions.

The build_full_margin function could possibly be made more efficient by directly iterating over existing edges as opposed to rebuilding the entire product automaton.

When using only one TS state model (2d region pose for example), TS state received by the LTL planner (planner.py of ltl_automaton_planner) is not recognized as either part of the TS or as previous state. It probably as something to do with the variable format.

The 2D regions generation script seems to be connecting the last cell of a row with the first cell of the following row.

The planner has a parameter for getting the initial state from the agent instead of the TS config file. Implementation of this feature has not been completed yet.

The package relies on NetworkX for graphs. Similar libraries, like Igraph or graph-tool, offer dramatically better perfomances (see https://www.timlrx.com/blog/benchmark-of-popular-graph-network-packages).

Therefor, one possible improvement could be to update the code with one of these libraries.

In some cases it could be useful to limit actions from one TS to only a few states of another TS. For example a pick/drop action could only be carried out in some 2D regions.

Previously, running the nexus e.g. irl_test launch file allowed to publish ts_state messages in order to change the TS state. The initial state is r1, unloaded, and a transition to r2, unloaded should be possible. Nevertheless, the error "Received TS states don't match TS state models: 2 initial states and 1 state models" is raised.

In a simple scenario with only 2D regions, all states are considered trap when testing the HIL

Commit d052b6e reformatted the TS state format to work with more than 2 dimensions. Even with one dimension in the TS, the TS state is now a tuple.

This seems to create an issue with one dimensional TS states, as they are not recognized as being part of the plan when received by the planner from the agent node.