diff --git a/conf/ppo/task/g1_motion_tracking/motrix.yaml b/conf/ppo/task/g1_motion_tracking/motrix.yaml
index 2a3343635..1b742d21f 100644
--- a/conf/ppo/task/g1_motion_tracking/motrix.yaml
+++ b/conf/ppo/task/g1_motion_tracking/motrix.yaml
@@ -38,6 +38,7 @@ reward:
     motion_joint_vel: 0.0
     action_rate_l2: -0.05
     joint_limit: -10.0
+    undesired_contacts: -0.1
   std_root_pos: 0.3
   std_root_ori: 0.4
   std_body_pos: 0.3
diff --git a/conf/ppo/task/g1_motion_tracking/mujoco.yaml b/conf/ppo/task/g1_motion_tracking/mujoco.yaml
index f257cd9e6..e53576b3a 100644
--- a/conf/ppo/task/g1_motion_tracking/mujoco.yaml
+++ b/conf/ppo/task/g1_motion_tracking/mujoco.yaml
@@ -24,6 +24,7 @@ reward:
     motion_joint_vel: 0.0
     action_rate_l2: -0.1
     joint_limit: -10.0
+    undesired_contacts: -0.1
   std_root_pos: 0.3
   std_root_ori: 0.4
   std_body_pos: 0.3
diff --git a/src/unilab/assets/motions/g1/run1_subject5_from_csv.npz b/src/unilab/assets/motions/g1/run1_subject5_from_csv.npz
new file mode 100644
index 000000000..fa8fa8fa2
Binary files /dev/null and b/src/unilab/assets/motions/g1/run1_subject5_from_csv.npz differ
diff --git a/src/unilab/envs/locomotion/g1/base.py b/src/unilab/envs/locomotion/g1/base.py
index f72cb419a..3f391c02f 100644
--- a/src/unilab/envs/locomotion/g1/base.py
+++ b/src/unilab/envs/locomotion/g1/base.py
@@ -13,10 +13,10 @@
 
 @dataclass
 class NoiseConfig:
-    level: float = 0.0
-    scale_joint_angle: float = 0.02
-    scale_joint_vel: float = 0.3
-    scale_gyro: float = 0.1
+    level: float = 1.0
+    scale_joint_angle: float = 0.01
+    scale_joint_vel: float = 1.5
+    scale_gyro: float = 0.2
     scale_gravity: float = 0.05
     scale_linvel: float = 0.1
 
@@ -46,3 +46,14 @@ class G1BaseEnv(LocomotionBaseEnv):
     _cfg: G1BaseCfg
     _keyframe_name = "stand"
     _use_global_dtype = False
+
+    def _obs_noise(self, data: np.ndarray, scale: float) -> np.ndarray:
+        """Apply per-step uniform observation noise scaled by ``noise_config.level``."""
+        noise_cfg = self._cfg.noise_config
+        if noise_cfg.level > 0.0:
+            return data + (
+                np.random.uniform(-1.0, 1.0, data.shape).astype(data.dtype)
+                * noise_cfg.level
+                * scale
+            )
+        return data
diff --git a/src/unilab/envs/locomotion/g1/joystick.py b/src/unilab/envs/locomotion/g1/joystick.py
index 3aa7d9d01..948358333 100644
--- a/src/unilab/envs/locomotion/g1/joystick.py
+++ b/src/unilab/envs/locomotion/g1/joystick.py
@@ -217,7 +217,13 @@ def update_state(self, state: NpEnvState) -> NpEnvState:
     def _compute_obs(
         self, info: dict, linvel, gyro, gravity, dof_pos, dof_vel
     ) -> dict[str, np.ndarray]:
+        noise_cfg = self._cfg.noise_config
         diff = dof_pos - self.default_angles
+        gyro = self._obs_noise(gyro, noise_cfg.scale_gyro)
+        gravity = self._obs_noise(gravity, noise_cfg.scale_gravity)
+        diff = self._obs_noise(diff, noise_cfg.scale_joint_angle)
+        dof_vel = self._obs_noise(dof_vel, noise_cfg.scale_joint_vel)
+        linvel = self._obs_noise(linvel, noise_cfg.scale_linvel)
         command = info["commands"]
         last_actions = info.get("current_actions", np.zeros_like(diff))
         gait_phase = info.get("gait_phase", np.zeros((self._num_envs, 2), dtype=get_global_dtype()))
diff --git a/src/unilab/envs/locomotion/g1/joystick_sac.py b/src/unilab/envs/locomotion/g1/joystick_sac.py
index 3699a60a9..3759e77a5 100644
--- a/src/unilab/envs/locomotion/g1/joystick_sac.py
+++ b/src/unilab/envs/locomotion/g1/joystick_sac.py
@@ -101,7 +101,13 @@ def __init__(self, cfg: G1JoystickSACCfg, num_envs=1, backend_type="mujoco"):
     def _compute_obs(
         self, info: dict, linvel, gyro, gravity, dof_pos, dof_vel
     ) -> dict[str, np.ndarray]:
+        noise_cfg = self._cfg.noise_config
         diff = dof_pos - self.default_angles
+        gyro = self._obs_noise(gyro, noise_cfg.scale_gyro)
+        gravity = self._obs_noise(gravity, noise_cfg.scale_gravity)
+        diff = self._obs_noise(diff, noise_cfg.scale_joint_angle)
+        dof_vel = self._obs_noise(dof_vel, noise_cfg.scale_joint_vel)
+        linvel = self._obs_noise(linvel, noise_cfg.scale_linvel)
         command = info["commands"]
         last_actions = info.get("current_actions", np.zeros_like(diff))
         gait_phase = info.get("gait_phase", np.zeros((self._num_envs, 2), dtype=get_global_dtype()))
diff --git a/src/unilab/envs/motion_tracking/g1/tracking.py b/src/unilab/envs/motion_tracking/g1/tracking.py
index 3225b08bf..90053fbd8 100644
--- a/src/unilab/envs/motion_tracking/g1/tracking.py
+++ b/src/unilab/envs/motion_tracking/g1/tracking.py
@@ -57,6 +57,7 @@ class RewardConfig:
             "motion_joint_vel": 0.0,
             "action_rate_l2": -0.1,
             "joint_limit": -10.0,
+            "undesired_contacts": -0.1,
         }
     )
     # Standard deviations for exponential rewards
@@ -119,8 +120,8 @@ class G1MotionTrackingCfg(G1BaseCfg):
         ASSETS_ROOT_PATH / "motions" / "g1" / "dance1_subject2_part.npz"
     )
     # motion_file: str | list[str] = str(ASSETS_ROOT_PATH / "motions" / "g1" / "gangnam_style.npz")
-    # motion_file: str | list[str] = str(ASSETS_ROOT_PATH / "motions" / "g1" / "walk1_subject5_from_csv.npz") #LAFAN
-    # motion_file: str | list[str] = str(ASSETS_ROOT_PATH / "motions" / "g1" / "sprint1_subject4_from_csv.npz") #LAFAN
+    # motion_file: str | list[str] = str(ASSETS_ROOT_PATH / "motions" / "g1" / "fight1_subject5_from_csv.npz") #LAFAN
+    # motion_file: str | list[str] = str(ASSETS_ROOT_PATH / "motions" / "g1" / "dance_basic_slide_180_R_loop_001__A322_M.npz") #LAFAN
     # motion_file: str | list[str] = str(ASSETS_ROOT_PATH / "motions" / "g1" / "playing_violin_R_003__A327_from_csv.npz") #Seed
     anchor_body_name: str = "torso_link"
     body_names: tuple[str, ...] = (
@@ -156,6 +157,7 @@ class G1MotionTrackingCfg(G1BaseCfg):
         "left_wrist_yaw_link",
         "right_wrist_yaw_link",
     )
+    undesired_contact_z_threshold: float = 0.05
 
 
 @registry.envcfg("G1MotionTracking")
@@ -316,6 +318,13 @@ def __init__(self, cfg: G1MotionTrackingCfg, num_envs=1, backend_type="mujoco"):
             [cfg.body_names.index(name) for name in cfg.ee_body_names], dtype=np.int32
         )
 
+        # Get non-EE body indices for undesired contact penalty
+        ee_set = set(cfg.ee_body_names)
+        self.undesired_contact_body_indices = np.array(
+            [i for i, name in enumerate(cfg.body_names) if name not in ee_set],
+            dtype=np.int32,
+        )
+
         # Load motion data
         self.motion_loader = MotionLoader(cfg.motion_file, body_indices=motion_body_ids)
         self.motion_sampler = MotionSampler(
@@ -366,6 +375,7 @@ def _init_reward_functions(self):
             "motion_joint_vel": self._reward_motion_joint_vel,
             "action_rate_l2": self._reward_action_rate_l2,
             "joint_limit": self._reward_joint_limit,
+            "undesired_contacts": self._reward_undesired_contacts,
         }
 
     def update_state(self, state: NpEnvState) -> NpEnvState:
@@ -543,6 +553,13 @@ def _compute_obs(
         joint_pos_rel = dof_pos - self.default_angles
         last_actions = info.get("current_actions", np.zeros_like(joint_pos_rel))
 
+        # Per-step observation noise on sensor channels
+        noise_cfg = self._cfg.noise_config
+        linvel = self._obs_noise(linvel, noise_cfg.scale_linvel)
+        gyro = self._obs_noise(gyro, noise_cfg.scale_gyro)
+        joint_pos_rel = self._obs_noise(joint_pos_rel, noise_cfg.scale_joint_angle)
+        dof_vel = self._obs_noise(dof_vel, noise_cfg.scale_joint_vel)
+
         # Actor observations
         actor_obs = np.concatenate(
             [
@@ -703,6 +720,12 @@ def _reward_motion_joint_vel(self, info: dict) -> np.ndarray:
             np.exp(-error / self._cfg.reward_config.std_joint_vel**2), dtype=get_global_dtype()
         )
 
+    def _reward_undesired_contacts(self, info: dict) -> np.ndarray:
+        robot_body_pos_w = info["robot_body_pos_w"]
+        body_z = robot_body_pos_w[:, self.undesired_contact_body_indices, 2]
+        is_contact = body_z < self._cfg.undesired_contact_z_threshold
+        return np.asarray(is_contact.sum(axis=-1), dtype=get_global_dtype())
+
     def _reward_action_rate_l2(self, info: dict) -> np.ndarray:
         action_diff = info["current_actions"] - info["last_actions"]
         return np.asarray(np.sum(np.square(action_diff), axis=1), dtype=get_global_dtype())
diff --git a/tests/envs/test_g1_obs_noise.py b/tests/envs/test_g1_obs_noise.py
new file mode 100644
index 000000000..77498eb17
--- /dev/null
+++ b/tests/envs/test_g1_obs_noise.py
@@ -0,0 +1,78 @@
+"""Tests for G1 per-step observation noise."""
+
+from __future__ import annotations
+
+import numpy as np
+
+from unilab.envs.locomotion.g1.base import G1BaseCfg, G1BaseEnv, NoiseConfig
+
+
+class _ConcreteG1Env(G1BaseEnv):
+    """Minimal concrete subclass — only needed to satisfy the ABC."""
+
+    def update_state(self, state):
+        raise NotImplementedError
+
+
+def _make_env(level: float) -> G1BaseEnv:
+    cfg = G1BaseCfg(noise_config=NoiseConfig(level=level))
+    env = object.__new__(_ConcreteG1Env)
+    env._cfg = cfg
+    return env
+
+
+class TestObsNoise:
+    def test_noise_applied_when_level_positive(self):
+        env = _make_env(level=1.0)
+        data = np.ones((4, 10), dtype=np.float32)
+        cfg = env._cfg.noise_config
+
+        results = [env._obs_noise(data.copy(), cfg.scale_joint_angle) for _ in range(5)]
+        # At least one result should differ from the original
+        assert any(not np.allclose(r, data) for r in results)
+
+    def test_no_noise_when_level_zero(self):
+        env = _make_env(level=0.0)
+        data = np.ones((4, 10), dtype=np.float32)
+        cfg = env._cfg.noise_config
+
+        result = env._obs_noise(data.copy(), cfg.scale_joint_angle)
+        np.testing.assert_array_equal(result, data)
+
+    def test_noise_bounded_by_level_times_scale(self):
+        env = _make_env(level=1.0)
+        data = np.zeros((128, 29), dtype=np.float32)
+        scale = 0.2
+        result = env._obs_noise(data.copy(), scale)
+        # uniform[-1,1] * 1.0 * 0.2 => bounded by [-0.2, 0.2]
+        assert np.all(result >= -scale)
+        assert np.all(result <= scale)
+
+    def test_noise_scales_with_level(self):
+        env_half = _make_env(level=0.5)
+        env_full = _make_env(level=1.0)
+        np.random.seed(42)
+        data = np.zeros((1024, 10), dtype=np.float32)
+        scale = 1.0
+
+        np.random.seed(0)
+        r_half = env_half._obs_noise(data.copy(), scale)
+        np.random.seed(0)
+        r_full = env_full._obs_noise(data.copy(), scale)
+
+        # Same random seed: full-level noise should be exactly 2x half-level noise
+        np.testing.assert_allclose(r_full, r_half * 2.0)
+
+    def test_noise_preserves_dtype(self):
+        for dt in [np.float32, np.float64]:
+            env = _make_env(level=1.0)
+            data = np.ones((4, 5), dtype=dt)
+            result = env._obs_noise(data, 0.1)
+            assert result.dtype == dt
+
+    def test_noise_preserves_shape(self):
+        env = _make_env(level=1.0)
+        for shape in [(1, 3), (64, 29), (1024, 10)]:
+            data = np.zeros(shape, dtype=np.float32)
+            result = env._obs_noise(data, 0.1)
+            assert result.shape == shape