A framework for automatically translating Numeric PDDL domains with the Gymnasium API. This allows Numeric PDDL planning problems to be solved using standard RL algorithms (e.g., PPO via RLlib).
The environment converts states, actions, and goals from Numeric PDDL into fixed-size numeric vectors, enabling direct integration with deep RL libraries.
- Supports Numeric PDDL domains and problems
- Automatic grounding and vectorization of:
- Predicates
- Numeric fluents
- Goal conditions
- Compatible with the Gymnasium API
- Designed for RLlib (Ray) integration
- Works with standard deep RL algorithms (e.g., PPO)
The environment is configured via a dictionary (env_config).
| Parameter | Type | Relevant Env | Description |
|---|---|---|---|
domain_path |
Path |
All | Path to the PDDL domain file |
problems_list |
List[Path] |
All | List of problem files |
max_steps |
int |
All | Max steps per episode |
executing_algorithm |
str |
All | Name of RL algorithm (e.g., "PPO") |
masking_strategy |
"pre" or "post" |
PDDLMaskedEnv | "pre": Filters invalid actions before execution using PDDL checks."post": Learns invalid actions after execution from feedback. |
count_inapplicable |
bool |
PDDLMaskedEnv | Relevant when using "post" masking.If true, inapplicable actions count toward the episode length; if false, they are ignored. |
map_size |
int |
MinecraftEnv | Grid size, e.g., 6, 10, 15. |
pip install numeric_pddl_gym
With RL (RLlib + Torch):
pip install numeric_pddl_gym[rl_agents]
python rl_agents/ppo_pddl_rllib_agent.py
See example for a planning-based agent in:
test/test_fixed_script_agent.py
- Can't encode complex goal conditions.
- Agents must be retrained if the problem has a different number of fluents, predicates, or goal conditions.
- Currently does not support manual interaction with the environment.
- Designed for fixed-structure problems (no variable-sized domains).
- Enabling action applicability checking leads to slow runtime in large-scale problems.
Coming soon