diff --git a/src/lerobot/robots/unitree_g1/controllers/sonic_whole_body.py b/src/lerobot/robots/unitree_g1/controllers/sonic_whole_body.py
index bf96eaf51..7e84183f6 100644
--- a/src/lerobot/robots/unitree_g1/controllers/sonic_whole_body.py
+++ b/src/lerobot/robots/unitree_g1/controllers/sonic_whole_body.py
@@ -43,6 +43,7 @@ from ..g1_utils import (
     MUJOCO_TO_ISAACLAB,
     G1_29_JointIndex,
     get_gravity_orientation,
+    make_ort_session_options,
 )
 
 logger = logging.getLogger(__name__)
@@ -87,7 +88,12 @@ def load_policy(
     logger.info(f"Loading {policy_type.upper()} SONIC decoder from: {repo_id}/{filename}")
     decoder_path = hf_hub_download(repo_id=repo_id, filename=filename)
 
-    decoder = ort.InferenceSession(decoder_path)
+    # Cap the thread pool: onboard, this decoder steps at 50 Hz in the same process as the
+    # ZMQ camera server (capture + JPEG encode). Default session options let ORT grab every
+    # core on the NX, which starves the camera thread (stale frames) and jitters the control
+    # loop (limping gait). It is a small MLP, so 1 thread is enough and lowest-latency.
+    session_options = make_ort_session_options(intra_op_num_threads=1, inter_op_num_threads=1)
+    decoder = ort.InferenceSession(decoder_path, sess_options=session_options)
     logger.info(f"Decoder loaded: {decoder.get_inputs()[0].shape} → {decoder.get_outputs()[0].shape}")
 
     # Extract deploy constants from ONNX metadata
diff --git a/src/lerobot/robots/unitree_g1/g1_utils.py b/src/lerobot/robots/unitree_g1/g1_utils.py
index 5d1d4a018..f4bf92d0b 100644
--- a/src/lerobot/robots/unitree_g1/g1_utils.py
+++ b/src/lerobot/robots/unitree_g1/g1_utils.py
@@ -80,6 +80,26 @@ def get_gravity_orientation(quaternion: list[float] | np.ndarray) -> np.ndarray:
     return gravity_orientation
 
 
+def make_ort_session_options(
+    intra_op_num_threads: int | None = None, inter_op_num_threads: int | None = None
+):
+    """Build quiet ONNX Runtime SessionOptions, optionally capping the CPU thread pool.
+
+    These tiny MLP policies are latency-bound, not throughput-bound, so letting ORT grab
+    every core starves the real-time control loop / torch policy and causes stutter. Pass
+    1 intra + 1 inter thread for lowest-latency per-step inference.
+    """
+    import onnxruntime as ort
+
+    so = ort.SessionOptions()
+    so.log_severity_level = 3
+    if intra_op_num_threads is not None:
+        so.intra_op_num_threads = intra_op_num_threads
+    if inter_op_num_threads is not None:
+        so.inter_op_num_threads = inter_op_num_threads
+    return so
+
+
 class G1_29_JointArmIndex(IntEnum):
     # Left arm
     kLeftShoulderPitch = 15