direct and residual movement inconsistency fixed

ml/run_eval.py

@@ -33,11 +33,12 @@ def main():
         use_gui=args.gui,
         random_command=False,
         max_episode_steps=args.max_steps_per_episode,
+        robot_mode="direct"
     )
 
     # Multi-Phase Configurations Suite
     phase_configs = [
-        (CurriculumPhase.STAND_ONLY,          "STAND",                 np.array([0.0, 0.0, 0.0], dtype=np.float32)),
+        #(CurriculumPhase.STAND_ONLY,          "STAND",                 np.array([0.0, 0.0, 0.0], dtype=np.float32)),
         (CurriculumPhase.FORWARD,             "FORWARD GAIT",           np.array([1.0, 0.0, 0.0], dtype=np.float32)),
         (CurriculumPhase.TURN_AND_DIRECTION,  "FORWARD + YAW TURN",     np.array([0.5, 0.0, 0.4], dtype=np.float32)),
         (CurriculumPhase.OMNI_DIRECTION,      "STRIDE LATERAL",         np.array([0.5, 0.5, 0.0], dtype=np.float32)),
@@ -62,9 +63,18 @@ def main():
             for ep in range(args.episodes_per_phase):
                 obs, _ = env.reset()
 
-                # Force environment into active curriculum phase and lock command
+                # Force environment into active curriculum phase and lock command BEFORE getting obs
                 env.curriculum_phase = phase_enum
                 env.command = test_cmd.copy()
+                cmd_vx, cmd_vy, cmd_omega = test_cmd
+                env.robot.robot_state = (
+                    "walking"
+                    if (abs(cmd_vx) > 0.01 or abs(cmd_vy) > 0.01 or abs(cmd_omega) > 0.01)
+                    else "idle"
+                )
+                env.robot.vector_dirmov = [float(cmd_vx), float(cmd_vy), float(cmd_omega)]
+
+                # Get correct observation with test_cmd attached
                 obs = env._get_obs()
 
                 done = False
@@ -72,12 +82,10 @@ def main():
                 steps = 0
 
                 while not done:
-                    # Enforce locked command each step
-                    env.command = test_cmd.copy()
-
                     # Predict deterministic action from policy
                     action, _ = model.predict(obs, deterministic=True)
                     
+                    # Step environment
                     obs, reward, terminated, truncated, _ = env.step(action)
                     done = terminated or truncated