perf(unitree_g1): cap ONNX Runtime thread pool for locomotion policies

Add make_ort_session_options() to g1_utils and use it in the GR00T locomotion
controller. These small MLP policies are latency-bound and run at 50 Hz in a
background thread alongside the torch upper-body policy and IK; letting ORT grab
every core starves the real-time loop and causes sim stutter, so cap it to 1
intra/inter thread for lowest-latency per-step inference.

Co-authored-by: Cursor <cursoragent@cursor.com>

src/lerobot/robots/unitree_g1/controllers/gr00t_locomotion.py

@@ -26,6 +26,7 @@ from ..g1_utils import (
     REMOTE_BUTTONS,
     G1_29_JointIndex,
     get_gravity_orientation,
+    make_ort_session_options,
 )
 
 logger = logging.getLogger(__name__)
@@ -68,9 +69,13 @@ def load_groot_policies(
         filename="GR00T-WholeBodyControl-Walk.onnx",
     )
 
-    # Load ONNX policies
+    # Load ONNX policies with a capped thread pool. GR00T runs at 50 Hz in a
-    policy_balance = ort.InferenceSession(balance_path)
+    # background thread alongside the (torch) upper-body policy and IK; letting ORT
-    policy_walk = ort.InferenceSession(walk_path)
+    # grab every core starves those and makes the whole rollout stutter. These are
+    # small MLPs, so 1 thread is both enough and lowest-latency.
+    so = make_ort_session_options(intra_op_num_threads=1, inter_op_num_threads=1)
+    policy_balance = ort.InferenceSession(balance_path, sess_options=so)
+    policy_walk = ort.InferenceSession(walk_path, sess_options=so)
 
     logger.info("GR00T policies loaded successfully")
 

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