diff --git a/selfdrive/controls/lib/latcontrol_torque.py b/selfdrive/controls/lib/latcontrol_torque.py
index dffe85c473..5a2814e089 100644
--- a/selfdrive/controls/lib/latcontrol_torque.py
+++ b/selfdrive/controls/lib/latcontrol_torque.py
@@ -52,10 +52,8 @@ class LatControlTorque(LatControl):
       actual_curvature = -VM.calc_curvature(math.radians(CS.steeringAngleDeg - params.angleOffsetDeg), CS.vEgo, params.roll)
       roll_compensation = params.roll * ACCELERATION_DUE_TO_GRAVITY
       curvature_deadzone = abs(VM.calc_curvature(math.radians(self.steering_angle_deadzone_deg), CS.vEgo, 0.0))
-      desired_lateral_accel = desired_curvature * CS.vEgo ** 2
 
-      # desired rate is the desired rate of change in the setpoint, not the absolute desired curvature
-      # desired_lateral_jerk = desired_curvature_rate * CS.vEgo ** 2
+      desired_lateral_accel = desired_curvature * CS.vEgo ** 2
       actual_lateral_accel = actual_curvature * CS.vEgo ** 2
       lateral_accel_deadzone = curvature_deadzone * CS.vEgo ** 2
 
@@ -67,6 +65,8 @@ class LatControlTorque(LatControl):
       # do error correction in lateral acceleration space, convert at end to handle non-linear torque responses correctly
       pid_log.error = float(setpoint - measurement)
       ff = gravity_adjusted_lateral_accel
+      # latAccelOffset corrects roll compensation bias from device roll misalignment relative to car roll
+      ff -= self.torque_params.latAccelOffset
       ff += get_friction(desired_lateral_accel - actual_lateral_accel, lateral_accel_deadzone, FRICTION_THRESHOLD, self.torque_params)
 
       freeze_integrator = steer_limited_by_safety or CS.steeringPressed or CS.vEgo < 5
diff --git a/selfdrive/controls/tests/test_torqued_lat_accel_offset.py b/selfdrive/controls/tests/test_torqued_lat_accel_offset.py
new file mode 100644
index 0000000000..5ba9980020
--- /dev/null
+++ b/selfdrive/controls/tests/test_torqued_lat_accel_offset.py
@@ -0,0 +1,71 @@
+import numpy as np
+from cereal import car, messaging
+from opendbc.car import ACCELERATION_DUE_TO_GRAVITY
+from opendbc.car import structs
+from opendbc.car.lateral import get_friction, FRICTION_THRESHOLD
+from openpilot.common.realtime import DT_MDL
+from openpilot.selfdrive.locationd.torqued import TorqueEstimator, MIN_BUCKET_POINTS, POINTS_PER_BUCKET, STEER_BUCKET_BOUNDS
+
+np.random.seed(0)
+
+LA_ERR_STD = 1.0
+INPUT_NOISE_STD = 0.1
+V_EGO = 30.0
+
+WARMUP_BUCKET_POINTS = (1.5*MIN_BUCKET_POINTS).astype(int)
+STRAIGHT_ROAD_LA_BOUNDS = (0.02, 0.03)
+
+ROLL_BIAS_DEG = 1.0
+ROLL_COMPENSATION_BIAS = ACCELERATION_DUE_TO_GRAVITY*float(np.sin(np.deg2rad(ROLL_BIAS_DEG)))
+TORQUE_TUNE = structs.CarParams.LateralTorqueTuning(latAccelFactor=2.0, latAccelOffset=0.0, friction=0.2)
+TORQUE_TUNE_BIASED = structs.CarParams.LateralTorqueTuning(latAccelFactor=2.0, latAccelOffset=-ROLL_COMPENSATION_BIAS, friction=0.2)
+
+
+def generate_inputs(torque_tune, la_err_std, input_noise_std=None):
+  rng = np.random.default_rng(0)
+  steer_torques = np.concat([rng.uniform(bnd[0], bnd[1], pts) for bnd, pts in zip(STEER_BUCKET_BOUNDS, WARMUP_BUCKET_POINTS, strict=True)])
+  la_errs = rng.normal(scale=la_err_std, size=steer_torques.size)
+  frictions = np.array([get_friction(la_err, 0.0, FRICTION_THRESHOLD, torque_tune) for la_err in la_errs])
+  lat_accels = torque_tune.latAccelFactor*steer_torques + torque_tune.latAccelOffset + frictions
+  if input_noise_std is not None:
+    steer_torques += rng.normal(scale=input_noise_std, size=steer_torques.size)
+    lat_accels += rng.normal(scale=input_noise_std, size=steer_torques.size)
+  return steer_torques, lat_accels
+
+def get_warmed_up_estimator(steer_torques, lat_accels):
+  est = TorqueEstimator(car.CarParams())
+  for steer_torque, lat_accel in zip(steer_torques, lat_accels, strict=True):
+    est.filtered_points.add_point(steer_torque, lat_accel)
+  return est
+
+def simulate_straight_road_msgs(est):
+  carControl = messaging.new_message('carControl').carControl
+  carOutput = messaging.new_message('carOutput').carOutput
+  carState = messaging.new_message('carState').carState
+  livePose = messaging.new_message('livePose').livePose
+  carControl.latActive = True
+  carState.vEgo = V_EGO
+  carState.steeringPressed = False
+  ts = DT_MDL*np.arange(2*POINTS_PER_BUCKET)
+  steer_torques = np.concat((np.linspace(-0.03, -0.02, POINTS_PER_BUCKET), np.linspace(0.02, 0.03, POINTS_PER_BUCKET)))
+  lat_accels = TORQUE_TUNE.latAccelFactor * steer_torques
+  for t, steer_torque, lat_accel in zip(ts, steer_torques, lat_accels, strict=True):
+    carOutput.actuatorsOutput.torque = float(-steer_torque)
+    livePose.orientationNED.x = float(np.deg2rad(ROLL_BIAS_DEG))
+    livePose.angularVelocityDevice.z = float(lat_accel / V_EGO)
+    for which, msg in (('carControl', carControl), ('carOutput', carOutput), ('carState', carState), ('livePose', livePose)):
+      est.handle_log(t, which, msg)
+
+def test_estimated_offset():
+  steer_torques, lat_accels = generate_inputs(TORQUE_TUNE_BIASED, la_err_std=LA_ERR_STD, input_noise_std=INPUT_NOISE_STD)
+  est = get_warmed_up_estimator(steer_torques, lat_accels)
+  msg = est.get_msg()
+  # TODO add lataccelfactor and friction check when we have more accurate estimates
+  assert abs(msg.liveTorqueParameters.latAccelOffsetRaw - TORQUE_TUNE_BIASED.latAccelOffset) < 0.03
+
+def test_straight_road_roll_bias():
+  steer_torques, lat_accels = generate_inputs(TORQUE_TUNE, la_err_std=LA_ERR_STD, input_noise_std=INPUT_NOISE_STD)
+  est = get_warmed_up_estimator(steer_torques, lat_accels)
+  simulate_straight_road_msgs(est)
+  msg = est.get_msg()
+  assert (msg.liveTorqueParameters.latAccelOffsetRaw < -0.05) and np.isfinite(msg.liveTorqueParameters.latAccelOffsetRaw)
diff --git a/selfdrive/test/process_replay/ref_commit b/selfdrive/test/process_replay/ref_commit
index 8e4ff1e2a8..7d21526144 100644
--- a/selfdrive/test/process_replay/ref_commit
+++ b/selfdrive/test/process_replay/ref_commit
@@ -1 +1 @@
-209b47bea61e145cf2d27eb3ab650c97bcd1d33f
\ No newline at end of file
+866f2cb1f0e49b2cb7115aa8131164b3a75fb2c5
\ No newline at end of file