Fix action smoothing history and viewer timing control

motrix_envs/src/motrix_envs/locomotion/go1/walk_np.py

@@ -172,8 +172,10 @@ class Go1WalkTask(NpEnv):
 
     def apply_action(self, actions, state):
         state.info["last_dof_vel"] = self.get_dof_vel(state.data)
-        state.info["last_actions"] = state.info["current_actions"]
-        state.info["current_actions"] = actions
+        if "last_last_actions" in state.info:
+            state.info["last_last_actions"] = state.info["last_actions"].copy()
+        state.info["last_actions"] = state.info["current_actions"].copy()
+        state.info["current_actions"] = np.array(actions, copy=True)
         state.data.actuator_ctrls = self._compute_torques(actions, state.data)
         return state
 

motrix_envs/tests/test_dreamwaq_state_safety.py

@@ -1,6 +1,9 @@
+from types import SimpleNamespace
+
 import numpy as np
 
 from motrix_envs.locomotion.go1.dreamwaq import _sanitize_dof_pos
+from motrix_envs.locomotion.go1.walk_np import Go1WalkTask
 
 
 def test_sanitize_dof_pos_handles_zero_quaternion_with_nonfinite_joints():
@@ -17,3 +20,25 @@ def test_sanitize_dof_pos_handles_zero_quaternion_with_nonfinite_joints():
     np.testing.assert_allclose(np.linalg.norm(clean[:, 3:7], axis=1), 1.0)
     assert clean[0, 7] == 0.0
     assert clean[1, 8] == 0.0
+
+
+def test_apply_action_advances_three_frame_action_history():
+    task = Go1WalkTask.__new__(Go1WalkTask)
+    task.get_dof_vel = lambda data: np.ones((1, 12), dtype=np.float32)
+    task._compute_torques = lambda actions, data: actions
+    state = SimpleNamespace(
+        data=SimpleNamespace(actuator_ctrls=None),
+        info={
+            "current_actions": np.full((1, 12), 2.0, dtype=np.float32),
+            "last_actions": np.full((1, 12), 1.0, dtype=np.float32),
+            "last_last_actions": np.zeros((1, 12), dtype=np.float32),
+        },
+    )
+    actions = np.full((1, 12), 3.0, dtype=np.float32)
+
+    task.apply_action(actions, state)
+    actions.fill(9.0)
+
+    np.testing.assert_array_equal(state.info["last_last_actions"], 1.0)
+    np.testing.assert_array_equal(state.info["last_actions"], 2.0)
+    np.testing.assert_array_equal(state.info["current_actions"], 3.0)

scripts/dreamwaq_sim2sim_mujoco.py

@@ -201,6 +201,10 @@ def parse_args():
     )
     parser.add_argument("--timeout", type=float, default=30.0, help="validation timeout in simulation seconds")
     parser.add_argument("--log-interval", type=float, default=0.5, help="pose log interval in simulation seconds")
+    parser.add_argument(
+        "--realtime-factor", type=float, default=1.0,
+        help="viewer playback speed relative to wall time (default: 1.0)",
+    )
     parser.add_argument("--manual", action="store_true", help="start with zero velocity and use keyboard commands")
     parser.add_argument("--headless", action="store_true", help="run validation without viewer or real-time delay")
     parser.add_argument("--no-validation", action="store_true", help="do not stop with automatic PASS/FAIL")
@@ -217,6 +221,9 @@ def main():
     if args.headless and args.no_validation:
         print("[ERROR] --headless requires automatic validation", file=sys.stderr)
         return 2
+    if args.realtime_factor <= 0.0:
+        print("[ERROR] --realtime-factor must be positive", file=sys.stderr)
+        return 2
 
     terrain_map = {
         "flat": "scene_dreamwaq_flat.xml",
@@ -292,6 +299,7 @@ def main():
             "down={platform_end:.2f}->{down_end:.2f}m pass_x={pass_x:.2f}m".format(**course)
         )
     if not args.headless:
+        print(f"[Viewer] realtime_factor={args.realtime_factor:.2f}x")
         print("[CTRL] W/S forward/back, Q/E lateral, A/D yaw, Space stop, R reset, Esc quit")
 
     keyboard = None
@@ -423,7 +431,7 @@ def main():
 
             if view is not None:
                 view.sync()
-                deadline = wall_start + data.time
+                deadline = wall_start + data.time / args.realtime_factor
                 remaining = deadline - time.monotonic()
                 if remaining > 0:
                     time.sleep(remaining)

scripts/run_dreamwaq_sim2sim.sh

@@ -11,6 +11,7 @@ terrain="stairs_box"
 level="0"
 forward_speed="0.5"
 timeout="30"
+realtime_factor="1.0"
 export_only=false
 manual=false
 headless=false
@@ -31,6 +32,8 @@ Options:
   -l, --level N          Terrain difficulty level passed to sim2sim. Default: 0.
       --speed MPS        Autonomous forward command. Default: 0.5.
       --timeout SEC      Validation timeout in simulation seconds. Default: 30.
+      --realtime-factor X
+                         Viewer playback speed. Default: 1.0 (real time).
       --manual           Start at zero velocity and use keyboard commands.
       --headless         Run validation without a viewer or real-time delay.
       --no-validation    Run until the viewer closes instead of returning PASS/FAIL.
@@ -82,6 +85,11 @@ while (($#)); do
             timeout="$2"
             shift 2
             ;;
+        --realtime-factor)
+            (($# >= 2)) || die "$1 requires a number"
+            realtime_factor="$2"
+            shift 2
+            ;;
         --manual)
             manual=true
             shift
@@ -115,6 +123,8 @@ esac
 [[ "$level" =~ ^[0-9]+$ ]] || die "level must be a non-negative integer: ${level}"
 [[ "$forward_speed" =~ ^[0-9]+([.][0-9]+)?$ ]] || die "speed must be a non-negative number: ${forward_speed}"
 [[ "$timeout" =~ ^[0-9]+([.][0-9]+)?$ ]] || die "timeout must be a non-negative number: ${timeout}"
+[[ "$realtime_factor" =~ ^[0-9]+([.][0-9]+)?$ ]] || die "realtime-factor must be a positive number: ${realtime_factor}"
+[[ "$realtime_factor" != "0" && "$realtime_factor" != "0.0" ]] || die "realtime-factor must be positive"
 command -v uv >/dev/null 2>&1 || die "uv is not available in PATH"
 
 if [[ -z "$checkpoint" ]]; then
@@ -159,6 +169,7 @@ sim_args=(
     --level "$level"
     --forward-speed "$forward_speed"
     --timeout "$timeout"
+    --realtime-factor "$realtime_factor"
 )
 [[ "$manual" == true ]] && sim_args+=(--manual)
 [[ "$headless" == true ]] && sim_args+=(--headless)