diff --git a/.gitignore b/.gitignore
index 188d871..57d902a 100644
--- a/.gitignore
+++ b/.gitignore
@@ -2,3 +2,4 @@
 *.pyc
 .ipynb_checkpoints/
 __pycache__/
+data/
\ No newline at end of file
diff --git a/loader/data_loader.py b/loader/data_loader.py
new file mode 100644
index 0000000..c04cd96
--- /dev/null
+++ b/loader/data_loader.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+import logging
+from torch.utils.data import Dataset
+import os
+import scipy.io as sio
+#from skimage.transform import resize
+#from skimage.transform import rotate
+from PIL import Image
+import numpy as np
+from scipy import optimize
+
+np.set_printoptions(threshold=np.inf, suppress=True)
+import random
+import copy
+import math
+
+
+class OffroadLoader(Dataset):
+    def __init__(self, grid_size, train=True, demo=None, datadir='/root/medirl/data/', pre_train=False, tangent=False,
+                 more_kinematic=None):
+        assert grid_size % 2 == 0, "grid size must be even number"
+        self.grid_size = grid_size
+        if train:
+            self.data_dir = datadir + 'train_data/'
+        else:
+            self.data_dir = datadir + 'test_data/'
+
+        if demo is not None:
+            self.data_dir = datadir + '/irl_data/' + demo
+
+        items = os.listdir(self.data_dir)
+        self.data_list = []
+        for item in items:
+            self.data_list.append(self.data_dir + '/' + item)
+
+        #self.data_normalization = sio.loadmat(datadir + '/irl_data/train-data-mean-std.mat')
+        self.pre_train = pre_train
+
+        # kinematic related feature
+        self.center_idx = self.grid_size / 2
+    
+
+    def __getitem__(self, index):
+        data_mat = sio.loadmat(self.data_list[index])
+        feat, robot_state_feat, past_traj, future_traj = data_mat['feat'].copy(), data_mat['robot_state_data'], data_mat['past_traj'], data_mat['future_traj']
+        # visualize rgb
+        feat = np.vstack((feat, np.expand_dims(feat[2], axis=0)))
+        feat = np.vstack((feat, np.expand_dims(feat[3], axis=0)))
+        feat = np.vstack((feat, np.expand_dims(feat[4], axis=0)))
+        # normalize features locally
+        for i in range(5):
+            feat[i] = (feat[i] - np.mean(feat[i])) / np.std(feat[i])
+        # normalize robot state feature locally
+        robot_state_feat = (robot_state_feat - np.mean(robot_state_feat, axis=0, keepdims=True)) / np.std(robot_state_feat, axis=0, keepdims=True)
+
+ 
+        if self.pre_train:
+            target = data_mat['feat'][1].copy()  # copy the variance layer first
+            target[target < 0.5] = 0.0
+            target[target >= 0.5] = -1.0
+            return feat, target
+       
+        future_traj = self.auto_pad_future(future_traj[:, :2])
+        past_traj = self.auto_pad_past(past_traj[:, :2])
+
+        return feat, past_traj, future_traj, robot_state_feat
+
+    def __len__(self):
+        return len(self.data_list)
+
+    def auto_pad_past(self, traj):
+        """
+        add padding (NAN) to traj to keep traj length fixed.
+        traj shape needs to be fixed in order to use batch sampling
+        :param traj: numpy array. (traj_len, 2)
+        :return:
+        """
+        fixed_len = self.grid_size
+        if traj.shape[0] >= self.grid_size:
+            traj = traj[traj.shape[0]-self.grid_size:, :]
+            #raise ValueError('traj length {} must be less than grid_size {}'.format(traj.shape[0], self.grid_size))
+        pad_len = self.grid_size - traj.shape[0]
+        pad_array = np.full((pad_len, 2), np.nan)
+        output = np.vstack((traj, pad_array))
+        return output
+
+    def auto_pad_future(self, traj):
+        """
+        add padding (NAN) to traj to keep traj length fixed.
+        traj shape needs to be fixed in order to use batch sampling
+        :param traj: numpy array. (traj_len, 2)
+        :return:
+        """
+        fixed_len = self.grid_size
+        if traj.shape[0] >= self.grid_size:
+            traj = traj[:self.grid_size, :]
+            #raise ValueError('traj length {} must be less than grid_size {}'.format(traj.shape[0], self.grid_size))
+        pad_len = self.grid_size - traj.shape[0]
+        pad_array = np.full((pad_len, 2), np.nan)
+        output = np.vstack((traj, pad_array))
+        return output
\ No newline at end of file
diff --git a/loader/single_data_loader.py b/loader/single_data_loader.py
new file mode 100644
index 0000000..b866b07
--- /dev/null
+++ b/loader/single_data_loader.py
@@ -0,0 +1,82 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+import logging
+from torch.utils.data import Dataset
+import os
+import scipy.io as sio
+#from skimage.transform import resize
+#from skimage.transform import rotate
+from PIL import Image
+import numpy as np
+from scipy import optimize
+from IPython import embed
+np.set_printoptions(threshold=np.inf, suppress=True)
+import random
+import copy
+import math
+
+
+class OffroadLoader(Dataset):
+    def __init__(self, grid_size, train=True, demo=None, datadir='/root/medirl/data/', pre_train=False, tangent=False,
+                 more_kinematic=None):
+        assert grid_size % 2 == 0, "grid size must be even number"
+        self.grid_size = grid_size
+        self.image_fol = datadir + "demo_0"
+        #self.data_normalization = sio.loadmat(datadir + '/irl_data/train-data-mean-std.mat')
+        self.pre_train = pre_train
+
+        # kinematic related feature
+        self.center_idx = self.grid_size / 2
+    
+
+    def __getitem__(self, index):
+        goal_sink_feat = np.array(Image.open(self.image_fol+"/goal_sink.png"))
+        semantic_img_feat = np.array(Image.open(self.image_fol+"/semantic_img.png"))[:,:,0:3]
+        with open(self.image_fol+"/trajectory.npy", 'rb') as f:
+            traj = np.load(f)
+        # visualize rgb
+        feat = np.concatenate((goal_sink_feat, semantic_img_feat), axis = 2).T
+        # normalize features locally
+        for i in range(6):
+            feat[i] = (feat[i] - np.mean(feat[i])) / np.std(feat[i])
+        
+ 
+    
+
+        return feat, traj
+
+    def __len__(self):
+        return len(self.data_list)
+
+    def auto_pad_past(self, traj):
+        """
+        add padding (NAN) to traj to keep traj length fixed.
+        traj shape needs to be fixed in order to use batch sampling
+        :param traj: numpy array. (traj_len, 2)
+        :return:
+        """
+        fixed_len = self.grid_size
+        if traj.shape[0] >= self.grid_size:
+            traj = traj[traj.shape[0]-self.grid_size:, :]
+            #raise ValueError('traj length {} must be less than grid_size {}'.format(traj.shape[0], self.grid_size))
+        pad_len = self.grid_size - traj.shape[0]
+        pad_array = np.full((pad_len, 2), np.nan)
+        output = np.vstack((traj, pad_array))
+        return output
+
+    def auto_pad_future(self, traj):
+        """
+        add padding (NAN) to traj to keep traj length fixed.
+        traj shape needs to be fixed in order to use batch sampling
+        :param traj: numpy array. (traj_len, 2)
+        :return:
+        """
+        fixed_len = self.grid_size
+        if traj.shape[0] >= self.grid_size:
+            traj = traj[:self.grid_size, :]
+            #raise ValueError('traj length {} must be less than grid_size {}'.format(traj.shape[0], self.grid_size))
+        pad_len = self.grid_size - traj.shape[0]
+        pad_array = np.full((pad_len, 2), np.nan)
+        output = np.vstack((traj, pad_array))
+        return output
\ No newline at end of file
diff --git a/mdp/offroad_grid.py b/mdp/offroad_grid.py
index e8db667..99bd191 100644
--- a/mdp/offroad_grid.py
+++ b/mdp/offroad_grid.py
@@ -31,9 +31,9 @@ class OffroadGrid(object):
     """
 
     def __init__(self, grid_size, discount):
-        self.actions = ((0, 1), (1, 0), (0, -1), (-1, 0))  # 0->RIGHT, 1->UP, 2->LEFT, 3->DOWN
+        # self.actions = ((0, 1), (1, 0), (0, -1), (-1, 0))  # 0->RIGHT, 1->UP, 2->LEFT, 3->DOWN
         # 0->RIGHT, 1->RIGHT-UP, 2->UP, 3->LEFT-UP, 4->LEFT, 5->LEFT-DOWN, 6->DOWN, 7-RIGHT-DOWN
-        # self.actions = ((0, 1), (1, 1), (1, 0), (1, -1), (0, -1), (-1, -1), (-1, 0), (-1, 1))
+        self.actions = ((0, 1), (1, 1), (1, 0), (1, -1), (0, -1), (-1, -1), (-1, 0), (-1, 1))
         self.n_actions = len(self.actions)
         self.n_states = grid_size ** 2
         self.grid_size = grid_size
@@ -61,7 +61,7 @@ def load_feat(self, feat):
             feat_vector = feat[:, x, y]
             self.feat_mat.append(feat_vector)
 
-        self.feat_mat = np.stack(self.feat_mat, axis=0)
+        self.feat_mat = feat
 
     def find_feat_expect(self, traj):
         """
diff --git a/mdp/test_offroad_grid.py b/mdp/test_offroad_grid.py
index 14ddc1e..793cf16 100644
--- a/mdp/test_offroad_grid.py
+++ b/mdp/test_offroad_grid.py
@@ -1,10 +1,11 @@
 import unittest
 import visdom
-from irl.mdp.offroad_grid import OffroadGrid
+import sys
+sys.path.append("/root/medirl/")
+from mdp.offroad_grid import OffroadGrid
 import numpy as np
 import numpy.random as rn
-from loader.offroad_loader import OffroadLoader, get_max_min_feature, get_mean_feature
-
+from loader.single_data_loader import OffroadLoader
 
 def make_random_model():
     grid_size = rn.randint(2, 15)
@@ -18,11 +19,11 @@ def __init__(self, *args, **kwargs):
         self.vis = visdom.Visdom()
 
     def test_main(self):
-        grid_size = 30
+        grid_size = 32
         discount = 0.9
         model = OffroadGrid(grid_size, discount)
         loader = OffroadLoader(grid_size=grid_size)
-        feat, traj = loader[1]
+        feat, traj = loader[0]
         print(traj)
         traj[np.isnan(traj)] = 0.0  # replace nan as 0.0
         self.vis.heatmap(X=traj, opts=dict(colormap='Electric', title='height_mean'))
diff --git a/requirements.txt b/requirements.txt
index 5ef45bc..61bc785 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,9 +1,10 @@
 -- jupyter
 -- cython
--- numpy=1.15.4
+-- numpy==1.15.4
 -- numba==0.50.1
--- matplotlib=2.1.2
--- scipy=1.0.0
+-- matplotlib==2.1.2
+-- scipy==1.0.0
 -- seaborn==0.9.0
--- tqdm=4.19.4
--- visdom==0.1.8.3
\ No newline at end of file
+-- tqdm==4.19.4
+-- visdom==0.1.8.3
+-- scikit-image==0.14.1
\ No newline at end of file