TAMPEST

TAMPEST (Task and Motion Planning by Encoding into Satisfiability Testing) is a meta-framework for solving complex Task and Motion Planning (TAMP) problems. TAMP problems combine discrete task planning with low-level continuous motion planning and are fundamental for robotic autonomy in dynamic and partially known environments.

TAMPEST introduces a general open-source framework for modeling, solving, and benchmarking TAMP problems. It includes a specialization for incremental SMT-based planning via TemPEST, a temporal planner that serves as one of the back-end engines.

Requirements

The first requirement is Python3 since the framework is written in python:

apt-get install python3-dev python3-pip

For the motion planning part OMPL is needed:

wget https://ompl.kavrakilab.org/install-ompl-ubuntu.sh
chmod u+x install-ompl-ubuntu.sh
./install-ompl-ubuntu.sh --python

Other Python requirements:

pip3 install -r requirements.txt

TAMPEST relies on TemPEST for the SMT-based task planning capabilities.

pip3 install git+https://github.com/fbk-pso/tempest.git

Solver dependencies for TemPEST (see PySMT):

pysmt-install --z3

One of the task planners that can be used for numeric planning is ENHSP. It requires JAVA:

apt-get install openjdk-17-jdk

Installation

Install TAMPEST with pip:

pip3 install git+https://github.com/fbk-pso/tampest.git

Usage

TAMPEST is fully integrated with the Unified Planning framework. You must register the planner engines with the Unified Planning environment:

from unified_planning.shortcuts import *

# Register TAMPEST engines
env = get_environment()
env.factory.add_engine("tampest", "tampest.engine", "TampestEngine")
env.factory.add_meta_engine("tamp", "tampest.meta_engine", "TampMetaEngine")

problem = ...  # your task and motion planning problem

# Solve with TAMPEST engine (it support classical, numeric and temporal planning)
with OneshotPlanner(name="tampest") as planner:
    result = planner.solve(problem)
    print(result.plan)

# Solve with TAMPEST meta engine (it support classical and numeric planning)
with OneshotPlanner(name="tamp[enhsp]") as planner:
    result = planner.solve(problem)
    print(result.plan)

References

• E. Tosello, A. Valentini, A. Micheli. A Meta-Engine Framework for Interleaved Task and Motion Planning using Topological Refinements. ECAI 2024

• E. Tosello, A. Valentini, A. Micheli. Temporal Task and Motion Planning with Metric Time for Multiple Object Navigation. AAAI 2025

Reproducing Results from ECAI 2024

To reproduce the experimental results from the ECAI 2024 paper, use the following command-line interface:

python3 test.py --domain <domain-name> --tr <topological-refinement> --tp <task-planner-name> --mp <motion-planner> --dim <dim> --d <n> --c <m> [--capacity <robot_capacity>]

The possible domains are: doors, maze, delivery, rover.

The possible topological refinements are: none, unreach, obs, all.

The possible task planners are: tampest, tamer, enhsp, fast-downward.

The possible motion planners are: LazyRRT, RRT.

The possible dimensions are: 2D, 3D. 3D setups are available only for maze and rover.

The possible set of benchmark options are:

• doors: --d [1 2 4 6 8 10] --c [0 1 2 3], with d the number of doors to be open and c = {0 = [(0,0)], 1 = [(10,0)], 2 = [(0,10)], 3 = [(5,5)]} the number of extra configurations sampled within the initially reachable space, the unreachable space, or equally splitted between both;

• maze: --d [1 2 3 4 5 6 7 8 9 10] --c [0 1 2 3 4 5 6 7 8 9 10], with d the number of closed doors and c the number of locations to be visited;

• delivery: --d [1 2 4 6 8 10] --c [0 2 4 ... 48] --capacity [4 3 2 1], with d the number of closed doors, capacity the loading capacity of the robot, and c the configuration of the parcels. In detail, c = n = [nr, ng, dr, dg], where nr is the total number of red parcels, ng is the total number of green parcels, dr is the number of red parcels already at their delivery stations, and dg is the number of green parcels already at their delivery stations;

• rover: --d [2 4 6 8 10] --c [0 1 2 3 4], with d the number of samples (equally splitted between soils and rocks, each one separated from the robot by a closed door) and c the number of objectives to be photographed around the sample.

Refer to the paper for the detailed meaning of the domain parameters.

See here to reproduce the results related to up-pddl-stream.

Reproducing Results from AAAI 2025

To reproduce the experimental results from the AAAI 2025 paper, use the following command-line interface:

python3 test.py --domain <domain-name> --tr <topological-refinement> --r <r> --d <n> [--c <m>] [--n_pallets <n_pallets>] [--kit_size <kit_size> --n_kit <n_kit>] [--n_drivers <n_drivers>] [--n_tiles <n_tiles> --n_colors <n_colors>]

The possible domains are: tdoors, majsp, kitting, driverlog, floortile.

The possible topological refinements are: none, unreach, obs, all.

The possible set of benchmark options are:

• tdoors: --r [1 2 3] --d [1 2 4 6] --c [0 1 2 3], with r the number of robots, d the number of doors to be open and c = {0 = [(0,0)], 1 = [(10,0)], 2 = [(0,10)], 3 = [(5,5)]} the number of extra configurations sampled within the initially reachable space, the unreachable space, or equally splitted between both;

• majsp: --r [1 2 3] --d [1 2 4 6] --n_pallets [1 2 3], with r the number of robots, d the number of closed doors and n_pallets the number of pallets to treat.

• kitting: --r [1 2 3] --d [1 2 4 6] --kit_size [1 2 3] --n_kit [1 2 3], with r the number of robots, d the number of closed doors, kit_size the max size of the kits and n_kit the number of kits to collect.

• driverlog: --r [1 2 3] --d [1 2 4 6] --n_drivers [1 2 3], with r the number of robots, d the number of closed doors and n_drivers the number of available drivers.

• floortile: --r [1 2 3] --d [1 2 4 6] --n_tiles [1 2 4 6 8] --n_colors [1 2 3], with r the number of robots, d the number of closed doors, n_tiles the number of tiles to paint and n_colors the number of available colors.

Refer to the paper for the detailed meaning of the domain parameters.

License

TAMPEST is released under the GNU Lesser General Public License v3.0 (LGPL-3.0). See the LICENSE file for full details.

Contact

For questions, bug reports, or contributions, please open an issue on GitHub or contact the maintainers.