Make IDM use vehicle on ego path a reference

src/baselines/rule_policies.py

@@ -85,7 +85,7 @@ class IDMRulePolicy(BaseAlgorithm):
 
         # Default IDM parameters
         assert target_speed>0, 'negative target speed'
-        self.s_max = target_speed
+        self.v_max = target_speed
         self.a_max = np.array([3.]) # nominal acceleration
         self.tau = 0.5 # desired time headway
         self.b_pref = 2.5 # preferred deceleration
@@ -124,37 +124,89 @@ class IDMRulePolicy(BaseAlgorithm):
             action (np.ndarray): action for controlled agent to take
         """
         agent = self._env._agent
+        state = self._env._env.state.numpy()
         full_state = self._env._env.projected_state.numpy() #(nv, 5)
         ego_state = full_state[agent] # (5,)
-        s = ego_state[2]
-        xy = full_state[:,0:2] # (nv, 2)
+        v_ego = ego_state[2]
+        # xy = full_state[:,0:2] # (nv, 2)
         v = full_state[:,2:3] # (nv, 1)
-        psi = full_state[:,3:4] # (nv, 1)
+        # psi = full_state[:,3:4] # (nv, 1)
 
-        d, r, i = self.get_ego_dr(agent, xy, v, psi)
+        length = 20
+        step = 0.5
+        x, y = self._env._env._generate_paths(delta=step, n=length/step, is_distance=True)
+        heading = to_circle(np.arctan2(np.diff(y), np.diff(x)))
+        velocities = state[:,1]
 
-        # propagate environment forward at constant velocity
-        for t in self.t_future:
-            if t > 0:
-                xy2 = xy + t * v * np.vstack((np.cos(psi[:,0]), np.sin(psi[:,0]))).T
-                d2, r2, i2 = self.get_ego_dr(agent, xy2, v, psi)
+        # something like this could be done to also take future proximity of vehicles to ego path into account
+        # time_horizon = np.array(range(3))
+        # predictions = state[:,0:1] + np.outer(state[:,1], time_horizon)
 
-                # choose closer vehicle (now vs imagined)
-                if d2 < d:
-                    d, r, i = d2, r2, i2
+        paths = np.stack([x[:,:-1],y[:,:-1], heading], axis=1) # (nv x 3 x (path_length-1))
+        ego_path = paths[agent:agent+1] # (1 x 3 x path_length-1)
 
-        # Update environment interaction graph with i
-        if i:
-            self._env._env._graph._neighbor_dict={agent:[i]}
+        # (x,y,phi) of all vehicles
+        poses = np.expand_dims(full_state[:, [0,1,3]], 2) # (nv x 3 x 1)
 
-        if d == np.inf:
-            d_des = self.d_min
-        else:
-            d_des = self.d_min + self.tau * s + s * r / (2* (self.a_max*self.b_pref)**0.5 )
+        diff = ego_path - poses
+        diff[:, 2, :] = to_circle(diff[:, 2, :])
+
+        # Test if position and heading angle are close for some point on the future vehicle track
+        max_pos_error = 1
+        pos_close = np.sum(diff[:, 0:2, :]**2, 1) <= max_pos_error**2 # (nv x path_length-1)
+        max_deg_error = 20
+        heading_close = np.abs(diff[:, 2, :]) <= 20 * np.pi / 180 # (nv x path_length-1)
+        # For all vehicles get the path points where they are close to the ego path
+        close = np.logical_and(pos_close, heading_close) # (nv x path_length-1)
+        close[agent, :] = False # exclude ego agent
+
+        leader = agent
+        min_idx = np.Inf
+        # Determine vehicle that is closest to ego in terms of path coordinate
+        for veh_id in range(len(close)):
+            path_idx = np.nonzero(close[veh_id])[0]
+            # veh_id is never close to agent
+            if len(path_idx) == 0:
+                continue
+            # first path index where veh_id is close to agent
+            elif path_idx[0] < min_idx:
+                leader = veh_id
+                min_idx = path_idx[0]
+
+        # alternative vectorized code
+        # def findfirst(a):
+        #     idx = np.argwhere(a)
+        #     if len(idx) == 0:
+        #         return np.NaN
+        #     else:
+        #         return float(idx[0]) # float conversion, to get a numpy array of dtype=float64
+        # d = np.apply_along_axis(findfirst, 1, close) # (nv)
+        # if np.all(np.isnan(d)):
+        #     leader = agent
+        # else:
+        #     leader = np.nanargmin(d)
+        #     path_idx = d[leader]
+        #     min_idx = np.sqrt(np.sum(diff[leader, 0:2, path_idx]**2))
+
+
+        if leader != agent:
+            # distance along ego path to point with closest distance
+            d = step * min_idx
+            # add distance from ego path point with closest distance to actual vehicle position
+            d += np.sqrt(np.sum(diff[leader, 0:2, min_idx]**2))
+
+            # Update environment interaction graph with leader
+            self._env._env._graph._neighbor_dict={agent:[leader]}
+
+            delta_v = v_ego - v[leader, 0]
+            d_des = self.d_min + self.tau * v_ego + v_ego * delta_v / (2* (self.a_max*self.b_pref)**0.5 )
             d_des = max(d_des, self.d_min)
+        else:
+            d = np.Inf
+            d_des = self.d_min
 
         assert (d_des>= self.d_min)
-        action = self.a_max*(1 - (s/self.s_max)**4 - (d_des/d)**2)
+        action = self.a_max*(1 - (v_ego/self.v_max)**4 - (d_des/d)**2)
 
         # normalize action to range if env is a NormalizedActionSpace
         if isinstance(self._env, NormalizedActionSpace):
@@ -163,6 +215,7 @@ class IDMRulePolicy(BaseAlgorithm):
         assert action.shape==(1,)
         return action
 
+
     def get_ego_dr(self, agent:int, xy: np.ndarray,
         v: np.ndarray, psi: np.ndarray) -> Tuple[float, float, Optional[int]]:
         """