diff --git a/src/lerobot_robot_ufactory/scripts/uf_lerobot_eval.py b/src/lerobot_robot_ufactory/scripts/uf_lerobot_eval.py
index ee8772c..f698f20 100644
--- a/src/lerobot_robot_ufactory/scripts/uf_lerobot_eval.py
+++ b/src/lerobot_robot_ufactory/scripts/uf_lerobot_eval.py
@@ -38,6 +38,7 @@ from lerobot.configs.policies import PreTrainedConfig
 from lerobot.scripts.lerobot_record import DatasetRecordConfig
 from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot_robot_ufactory.utils.utils import init_keyboard_listener
+from lerobot_robot_ufactory.utils.action_safety import ActionSafetyConfig, ActionSafetyGuard
 from lerobot_robot_ufactory.devices.umi.vive_tracker.transformations import Transformations
 
 
@@ -120,9 +121,11 @@ class EvalConfig:
         return ["policy"]
 
 
-def eval_loop(cfg: EvalConfig, relative=False, rx_continuous=False):
+def eval_loop(cfg: EvalConfig, relative=False, rx_continuous=False, safety_guard: ActionSafetyGuard | None = None):
     init_logging()
     logging.info(pformat(asdict(cfg)))
+    if safety_guard is not None:
+        safety_guard.log_config()
 
     robot = make_robot_from_config(cfg.robot)
 
@@ -261,6 +264,8 @@ def eval_loop(cfg: EvalConfig, relative=False, rx_continuous=False):
             prev_robot_dict[key] = {'type': 1 if is_tcp else 0, 'pose': np.array(pose)}
             prev_action_dict[key] = {'type': 1 if is_tcp else 0, 'pose': np.array(pose)}
 
+        safety_halted = False
+
         while True:
             start_loop_t = time.perf_counter()
 
@@ -269,6 +274,12 @@ def eval_loop(cfg: EvalConfig, relative=False, rx_continuous=False):
                 print("\n********** Policy Eval Episode (Reset) **********")
                 break
 
+            # Safety halt: stop inference and stop sending actions,
+            # wait for the operator to press the right arrow key to resume
+            if safety_halted:
+                precise_sleep(sleep_time_s)
+                continue
+
             # Get robot observation
             obs = robot.get_observation()
 
@@ -398,6 +409,16 @@ def eval_loop(cfg: EvalConfig, relative=False, rx_continuous=False):
             #     gripper_raw = (future_gripper_norm + 1) / 2 * (_gripper_max - _gripper_min) + _gripper_min
             # robot_action_to_send['right.gripper.pos'] = 1.0 if gripper_raw > 0.4 else 0.0
 
+            # Safety check: any violation triggers an e-stop; the action is
+            # NOT sent and the loop waits for the operator to press right arrow
+            if safety_guard is not None:
+                violation = safety_guard.check(robot_action_to_send, curr_robot_dict, keys)
+                if violation is not None:
+                    safety_halted = True
+                    logging.error(f"*** SAFETY HALT *** {violation}")
+                    print(f"\n*** SAFETY HALT *** {violation}\nAction was NOT sent. Press right arrow (->) to reset and resume, ESC to exit.")
+                    continue
+
             robot.send_action(robot_action_to_send)
         
             dt_s = time.perf_counter() - start_loop_t
@@ -418,11 +439,15 @@ def main():
     parser = argparse.ArgumentParser(description='configuration args')
     parser.add_argument('--relative', action='store_true', help='is relative motion or not')
     parser.add_argument('--rx_continuous', action='store_true', help='rx continuous or not')
+    parser.add_argument('--enable_safety', action='store_true', help='enable action safety guard (thresholds in ActionSafetyConfig)')
     args, unknown = parser.parse_known_args()
     sys.argv = [sys.argv[0]] + unknown
     register_third_party_plugins()
     cfg = get_cfg()
-    eval_loop(cfg, args.relative, args.rx_continuous)
+    # Action safety guard: tune thresholds in ActionSafetyConfig directly.
+    # Any violation triggers an e-stop; press right arrow to resume.
+    safety_guard = ActionSafetyGuard(ActionSafetyConfig(enabled=args.enable_safety))
+    eval_loop(cfg, args.relative, args.rx_continuous, safety_guard)
 
 
 if __name__ == "__main__":
diff --git a/src/lerobot_robot_ufactory/utils/action_safety.py b/src/lerobot_robot_ufactory/utils/action_safety.py
new file mode 100644
index 0000000..7fe6b80
--- /dev/null
+++ b/src/lerobot_robot_ufactory/utils/action_safety.py
@@ -0,0 +1,130 @@
+"""Model-agnostic action safety checks.
+
+Validates the final action dict right before it is sent to the robot; any
+failed check triggers an emergency halt. Only relies on the action key
+naming convention: `[prefix.]pose.x/y/z/rx/ry/rz`, `*.gripper.pos`, so it
+works with any policy (ACT / DP / pi0 / custom backbones) and with both
+single-arm and multi-arm setups.
+"""
+
+import logging
+import math
+from dataclasses import dataclass
+
+import numpy as np
+
+from lerobot_robot_ufactory.devices.umi.vive_tracker.transformations import Transformations
+
+POSE_AXES = ("x", "y", "z", "rx", "ry", "rz")
+
+
+@dataclass
+class ActionSafetyConfig:
+    """Configuration for ActionSafetyGuard.
+
+    All limits are compared against the robot's current actual pose per
+    control step. `workspace_min`/`workspace_max` use the same unit as the
+    pose commands (e.g. mm) and must be provided together.
+    """
+
+    enabled: bool = True
+    max_step_mm: float = 25.0
+    max_rot_step: float = 0.35
+    workspace_min: list | None = None
+    workspace_max: list | None = None
+
+    def __post_init__(self):
+        if (self.workspace_min is None) != (self.workspace_max is None):
+            raise ValueError("workspace_min and workspace_max must be provided together")
+
+
+def _rotation_delta_norm(rot_prev, rot_curr) -> float:
+    """Relative rotation angle (rad) between two axis-angle rotations.
+
+    Uses rotation matrices for the diff to avoid the ±π discontinuity
+    of subtracting raw rotvecs.
+    """
+    R_prev = Transformations.rxryrz_to_rotation_matrix(*rot_prev)
+    R_curr = Transformations.rxryrz_to_rotation_matrix(*rot_curr)
+    R_delta = R_prev.T @ R_curr
+    delta = Transformations.rotation_matrix_to_rxryrz(R_delta)
+    return float(np.linalg.norm(delta))
+
+
+class ActionSafetyGuard:
+    """Safety checks for the action dict about to be sent to the robot.
+
+    check() returns None when the action is safe; otherwise it returns a
+    human-readable violation reason (the caller should trigger an e-stop).
+    """
+
+    def __init__(self, config: ActionSafetyConfig | None = None):
+        config = config or ActionSafetyConfig()
+        self.config = config
+        self.enabled = config.enabled
+        self.max_step_mm = config.max_step_mm
+        self.max_rot_step = config.max_rot_step
+        self.workspace_min = None if config.workspace_min is None else np.asarray(config.workspace_min, dtype=np.float64)
+        self.workspace_max = None if config.workspace_max is None else np.asarray(config.workspace_max, dtype=np.float64)
+
+    def check(self, action: dict, curr_robot_dict: dict, keys) -> str | None:
+        """Check the action dict. `action` holds absolute pose commands and
+        `curr_robot_dict` holds each arm's current actual pose."""
+        if not self.enabled:
+            return None
+
+        for key in keys:
+            prefix = f".{key}" if key else ""
+            pose_keys = [f"{prefix}pose.{axis}" for axis in POSE_AXES]
+            if not all(k in action for k in pose_keys):
+                continue  # not TCP-pose controlled (e.g. joint-only arm), skip pose checks
+
+            target = np.array([action[k] for k in pose_keys], dtype=np.float64)
+
+            # 1. Numeric validity: NaN / Inf
+            if not np.all(np.isfinite(target)):
+                return f"[{key or 'arm'}] action contains NaN/Inf: {target.tolist()}"
+
+            # 2. Single-step delta limit, measured from the robot's current actual pose
+            curr = curr_robot_dict.get(key)
+            if curr is not None and curr.get("type") == 1:
+                curr_pose = np.asarray(curr["pose"], dtype=np.float64)
+                pos_dist = float(np.linalg.norm(target[:3] - curr_pose[:3]))
+                if pos_dist > self.max_step_mm:
+                    return (
+                        f"[{key or 'arm'}] position step {pos_dist:.1f}mm exceeds limit "
+                        f"{self.max_step_mm}mm (current {curr_pose[:3].tolist()} -> target {target[:3].tolist()})"
+                    )
+                rot_delta = _rotation_delta_norm(curr_pose[3:6], target[3:6])
+                if rot_delta > self.max_rot_step:
+                    return (
+                        f"[{key or 'arm'}] rotation step {rot_delta:.3f}rad exceeds limit "
+                        f"{self.max_rot_step}rad"
+                    )
+
+            # 3. Workspace bounding box (optional)
+            if self.workspace_min is not None:
+                pos = target[:3]
+                if np.any(pos < self.workspace_min) or np.any(pos > self.workspace_max):
+                    return (
+                        f"[{key or 'arm'}] target position {pos.tolist()} outside workspace "
+                        f"[{self.workspace_min.tolist()}, {self.workspace_max.tolist()}]"
+                    )
+
+        # Gripper numeric check
+        for k, v in action.items():
+            if "gripper" in k and isinstance(v, (int, float)) and not math.isfinite(v):
+                return f"[{k}] gripper action contains NaN/Inf: {v}"
+
+        return None
+
+    def log_config(self):
+        ws = (
+            f"[{self.workspace_min.tolist()}, {self.workspace_max.tolist()}]"
+            if self.workspace_min is not None
+            else "not set"
+        )
+        logging.info(
+            f"ActionSafetyGuard: max_step={self.max_step_mm}mm, "
+            f"max_rot_step={self.max_rot_step}rad, workspace={ws}"
+        )