wip

scripts/kitchen_precompute_prm.py

@@ -93,7 +93,7 @@ def _main() -> None:
 
     # Sample random trajectories in 7 DOF space.
     noise_scale = 0.1
-    max_steps_per_traj = 100
+    max_steps_per_traj = 500
     num_trajs = 10
     noise = OrnsteinUhlenbeckActionNoise(np.zeros(7), sigma=noise_scale)
 
@@ -116,31 +116,45 @@ def _main() -> None:
             all_poses.append(pose)
             # TODO quit if something changes...
 
-    _reset_gym_env(gym_env)
-    init_pose = _get_pose_from_env(gym_env)
-    _add_pose_to_graph(init_pose, graph, distance_thresh)
-    reachable_nodes = set(nx.shortest_path(graph, init_pose))
+    obs = env.reset("train", 0)
+    state = env.state_info_to_state(obs["state_info"])
+    objs = set(state)
+    total_dist = 0.0
 
-    target_xyz = np.array([5.0, 0.0, 2.0])
-    target = min(reachable_nodes, key=lambda p: np.linalg.norm(p.xyz - target_xyz))
-    print("init:", init_pose)
-    print("target_xyz:", target_xyz)
-    print("target:", target)
+    for obj in objs:
+        if "gripper" in str(obj):
+            continue
+        print("TRYING TO REACH", obj)
+        obs = env.reset("train", 0)
+        state = env.state_info_to_state(obs["state_info"])
+        init_pose = _get_pose_from_env(env._gym_env)
+        x, y, z = state.get(obj, "x"), state.get(obj, "y"), state.get(obj, "z")
+        target_xyz = np.array([x, y, z])
+        graph_copy = graph.copy()
 
-    path = nx.shortest_path(graph, init_pose, target, weight="weight")
+        _add_pose_to_graph(init_pose, graph_copy, distance_thresh)
+        reachable_nodes = set(nx.shortest_path(graph_copy, init_pose))
 
-    for pose in path:
-        main_act = pose.joints
-        act = np.concatenate([main_act, gym_env.sim.data.qpos[7:9]])
-        gym_env.step(act)
-        gym_env.render()
+        target = min(reachable_nodes, key=lambda p: np.linalg.norm(p.xyz - target_xyz))
+        print("Closest node in graph distance:", np.linalg.norm(target.xyz - target_xyz))
+
+        path = nx.shortest_path(graph_copy, init_pose, target, weight="weight")
+
+        for pose in path:
+            main_act = pose.joints
+            act = np.concatenate([main_act, gym_env.sim.data.qpos[7:9]])
+            gym_env.step(act)
+            gym_env.render()
 
-    final_pose = _get_pose_from_env(gym_env)
-    print("Distance to target xyz:", np.linalg.norm(final_pose.xyz - target_xyz))
+        final_pose = _get_pose_from_env(gym_env)
+        dist = np.linalg.norm(final_pose.xyz - target_xyz)
+        print("Distance to target xyz:", dist)
+        total_dist += dist
+
+    print("TOTAL DIST:", total_dist)
 
     # TODO NEXT
-    # visualize important target locations in the environment, make sure they
-    # are reachable.
+    # check distance to important objects in the environment
     # incorporate collisions
     # incorporate orientation of end effector
     # sparsify graph