KEA Plan Generation

KEA Plan Generation provides utilities to help create PDDL domains and problems using LLMs. The package exposes two CLI tools:

• generate_prompt: create prompt datasets from a YAML config.
• train_llm: prototype RL-based training loop for plan generation.

Generate prompt datasets

The prompt generator expects a YAML config with fields like:

• prompt
• cell_grid_size (min_x, max_x, min_y, max_y)
• cell_contents
• num_robots
• goal_conditions
• robots_types
• action_capabilities
• action_time

Example:

ros2 run kea_plan_gen generate_prompt \
  --config /path/to/planner_constraints.yaml \
  --num_prompts 50 \
  --output prompts.txt

Train LLM (prototype)

The training script consumes a JSONL database where each line contains at least:

{"prompt": "..."}

Run:

ros2 run kea_plan_gen train_llm --db /path/to/prompts.jsonl --model_name <model>

Notes:

• The training loop currently uses a debug reward function and exits after one run.
• You will need additional Python dependencies (datasets, trl, peft).

This subsystem is intended as a starting point for building a plan-generation pipeline.