new way to load models with different action spaces

Author: ReykCS

agents/rto_tlabs/hyperparameters.json

@@ -1,24 +1,25 @@
 {
-    "agent_name": "AGENT_24_2022_02_04__11_36",
-    "robot": "rto_real",
-    "actions_in_observationspace": true,
-    "reward_fnc": "rule_05",
-    "discrete_action_space": false,
-    "normalize": false,
-    "task_mode": "staged",
-    "train_max_steps_per_episode": 120,
-    "eval_max_steps_per_episode": 170,
-    "goal_radius": 0.7,
-    "curr_stage": 10,
-    "batch_size": 19200,
-    "gamma": 0.99,
-    "n_steps": 2400,
-    "ent_coef": 0.005,
-    "learning_rate": 0.0003,
-    "vf_coef": 0.22,
-    "max_grad_norm": 0.5,
-    "gae_lambda": 0.95,
-    "m_batch_size": 15,
-    "n_epochs": 3,
-    "clip_range": 0.22
-}
+  "agent_name": "AGENT_24_2022_02_04__11_36",
+  "robot": "rto_real",
+  "actions_in_observationspace": true,
+  "reward_fnc": "rule_05",
+  "discrete_action_space": false,
+  "normalize": false,
+  "task_mode": "staged",
+  "train_max_steps_per_episode": 120,
+  "eval_max_steps_per_episode": 170,
+  "goal_radius": 0.7,
+  "curr_stage": 10,
+  "batch_size": 19200,
+  "gamma": 0.99,
+  "n_steps": 2400,
+  "ent_coef": 0.005,
+  "learning_rate": 0.0003,
+  "vf_coef": 0.22,
+  "max_grad_norm": 0.5,
+  "gae_lambda": 0.95,
+  "m_batch_size": 15,
+  "n_epochs": 3,
+  "clip_range": 0.22,
+  "observation_space": ["laser_scan", "goal_in_robot_frame", "last_action"]
+}

agents/rto_tlabs_marl/hyperparameters.json

@@ -1,24 +1,25 @@
 {
-    "agent_name": "AGENT_24_2022_02_04__11_36",
-    "robot": "rto_real",
-    "actions_in_observationspace": true,
-    "reward_fnc": "rule_06",
-    "discrete_action_space": false,
-    "normalize": false,
-    "task_mode": "staged",
-    "train_max_steps_per_episode": 400,
-    "eval_max_steps_per_episode": 1100,
-    "goal_radius": 0.7,
-    "curr_stage": 1,
-    "batch_size": 38400,
-    "gamma": 0.99,
-    "n_steps": 3200,
-    "ent_coef": 0.005,
-    "learning_rate": 0.0003,
-    "vf_coef": 0.22,
-    "max_grad_norm": 0.5,
-    "gae_lambda": 0.95,
-    "m_batch_size": 15,
-    "n_epochs": 3,
-    "clip_range": 0.22
-}
+  "agent_name": "AGENT_24_2022_02_04__11_36",
+  "robot": "rto_real",
+  "actions_in_observationspace": true,
+  "reward_fnc": "rule_06",
+  "discrete_action_space": false,
+  "normalize": false,
+  "task_mode": "staged",
+  "train_max_steps_per_episode": 400,
+  "eval_max_steps_per_episode": 1100,
+  "goal_radius": 0.7,
+  "curr_stage": 1,
+  "batch_size": 38400,
+  "gamma": 0.99,
+  "n_steps": 3200,
+  "ent_coef": 0.005,
+  "learning_rate": 0.0003,
+  "vf_coef": 0.22,
+  "max_grad_norm": 0.5,
+  "gae_lambda": 0.95,
+  "m_batch_size": 15,
+  "n_epochs": 3,
+  "clip_range": 0.22,
+  "observation_space": ["laser_scan", "goal_in_robot_frame", "last_action"]
+}

rosnav/utils/utils.py

@@ -11,11 +11,11 @@
 
 def get_robot_yaml_path(robot_model: str = None) -> str:
     if not robot_model:
-        robot_model = rospy.get_param("model")
+        robot_model = rospy.get_param("robot_model")
 
     simulation_setup_path = rospkg.RosPack().get_path("arena-simulation-setup")
     return os.path.join(
-        simulation_setup_path, "robot", robot_model, f"{robot_model}.model.yaml"
+        simulation_setup_path, "robot", robot_model, f"model_params.yaml"
     )
 
 
@@ -24,26 +24,20 @@ def get_laser_from_robot_yaml(robot_model: str = None) -> Tuple[int, int, int, i
 
     with open(robot_yaml_path, "r") as fd:
         robot_data = yaml.safe_load(fd)
+        laser_data = robot_data["laser"]
 
-        for plugin in robot_data["plugins"]:
-            if plugin["type"] == "Laser":
-                laser_angle_min = plugin["angle"]["min"]
-                laser_angle_max = plugin["angle"]["max"]
-                laser_angle_increment = plugin["angle"]["increment"]
+        rospy.set_param("laser/num_beams", laser_data["num_beams"])
 
-                _L = int(
-                    round((laser_angle_max - laser_angle_min) / laser_angle_increment)
-                )
-
-                # Because RosnavEncoder ist weird...
-                rospy.set_param("laser/num_beams", _L)
-
-                return _L, laser_angle_min, laser_angle_max, laser_angle_increment
+        return (
+            laser_data["num_beams"], 
+            laser_data["angle"]["min"], 
+            laser_data["angle"]["max"],
+            laser_data["angle"]["increment"]
+        )
 
 
 def get_observation_space_from_file(robot_model: str = None) -> Tuple[int, int]:
-    actions_in_obs = rospy.get_param("/actions_in_obs", True)
-    robot_state_size, action_state_size = 2, 3 if actions_in_obs else 0
+    robot_state_size, action_state_size = 2, rospy.get_param(rospy.get_namespace() + "action_state_size", 3)
     num_beams, _, _, _ = get_laser_from_robot_yaml(robot_model)
 
     return num_beams, action_state_size + robot_state_size

scripts/rosnav_node.py

@@ -2,8 +2,10 @@
 import rospkg
 import os
 import sys
+import traceback
 import json
 from stable_baselines3 import PPO
+from stable_baselines3.common.utils import get_device
 
 from rosnav.srv import GetAction, GetActionResponse
 from rosnav.rosnav_space_manager.rosnav_space_manager import RosnavSpaceManager
@@ -24,6 +26,7 @@ def __init__(self):
 
         # Load hyperparams
         self._hyperparams = self._load_hyperparams(self.agent_path)
+        # rospy.set_param("/actions_in_obs", self._hyperparams.get("actions_in_observationspace", False))
 
         self._obs_structure = self._get_observation_space_structure(self._hyperparams)
 
@@ -54,7 +57,16 @@ def _handle_next_action_srv(self, request):
         return response
 
     def _get_model(self, agent_path):
-        return PPO.load(os.path.join(agent_path, "best_model.zip")).policy
+        action_state_sizes = [0, 3]
+
+        for size in action_state_sizes:
+            rospy.set_param(rospy.get_namespace() + "action_state_size", size)
+            try:
+                return PPO.load(os.path.join(agent_path, "best_model.zip")).policy
+            except:
+                pass
+
+        rospy.signal_shutdown("")
 
     def _get_model_path(self, model_name):
         return os.path.join(