def generate_problems(num_train=50, num_test=10, level=1, **kwargs):
domain = PDDLDomainParser(os.path.join(PDDLDIR, f"{DOMAIN_NAME}.pddl"),
expect_action_preds=True,
operators_as_actions=False)
# Create version of the domain for simplicity
domain_name_with_level = f"{DOMAIN_NAME}_level{level}"
domain.write(os.path.join(PDDLDIR, f"{domain_name_with_level}.pddl"))
# Make sure problems are unique
seen_problem_ids = set()
problem_idx = 0
while problem_idx < num_train + num_test:
if problem_idx < num_train:
problem_dir = domain_name_with_level
else:
problem_dir = f"{domain_name_with_level}_test"
problem_outfile = f"problem{problem_idx}.pddl"
problem_id, problem_filepath = sample_problem(domain, problem_dir,
problem_outfile,
**kwargs)
if problem_id in seen_problem_ids:
continue
seen_problem_ids.add(problem_id)
if problem_is_valid(domain, problem_filepath):
problem_idx += 1
def run_probabilistic_planning_demo(env, planner_name, verbose=False, num_epi=20, outdir='/tmp', fps=3):
"""Probabilistic planning via simple determinization.
"""
if outdir is None:
outdir = "/tmp/{}".format(env_cls.__name__)
if not os.path.exists(outdir):
os.makedirs(outdir)
if env._render:
if env._problem_index_fixed:
problem_idx = env._problem_idx
video_path = os.path.join(outdir, 'planning_{}_{}_{}_demo.gif'.format(
planner_name, env.spec.id, problem_idx))
else:
video_path = os.path.join(outdir, 'planning_{}_{}_demo.gif'.format(
planner_name, env.spec.id))
env = VideoWrapper(env, video_path, fps=fps)
avg_reward = 0
for _ in range(num_epi):
obs, debug_info = env.reset()
domain = PDDLDomainParser(debug_info["domain_file"])
domain.determinize()
domain.write("/tmp/domain.pddl")
plan = run_planner("/tmp/domain.pddl", debug_info['problem_file'], planner_name)
actions = []
for s in plan:
a = parse_plan_step(
s,
env.domain.operators.values(),
env.action_predicates,
obs.objects,
operators_as_actions=env.operators_as_actions
)
actions.append(a)
tot_reward = 0.
for action in actions:
if verbose:
print("Obs:", obs)
if verbose:
print("Act:", action)
obs, reward, done, _ = env.step(action)
env.render()
tot_reward += reward
if verbose:
print("Rew:", reward)
if done:
break
if verbose:
print("Final obs:", obs)
print("Got total reward:", tot_reward)
print()
avg_reward += tot_reward/num_epi
print("Average reward over {} episodes was {}".format(num_epi, avg_reward))
env.close()
if verbose:
input("press enter to continue to next problem")
return tot_reward