Change tactile sensor normal component to report penetration depth.

The normal component of the tactile sensor now reports the maximum penetration depth at each taxel, instead of a derived normal force. The depth is negated so that positive values indicate penetration.

PiperOrigin-RevId: 921980899
Change-Id: Ide4103c0aff465e25a81cfdda650f6a0e2da9e0b

doc/XMLreference.rst

@@ -8514,15 +8514,14 @@ Extraction
    :width: 30%
    :target: https://github.com/google-deepmind/mujoco/blob/main/model/tactile/tactile.xml
 
-The tactile sensor returns the penetration pressure and the sliding velocities in the tangent frame at given points
-between the geom associated with the sensor and the SDF geoms in contact with it. We define the penetration pressure as
-a function of the penetration depth :math:`p(d) = \frac{d}{d_{max}-d}`, which is zero at the surface and goes to
-infinity as the maximum depth is reached. The sensor is associated with a geom and a mesh. It is activated by the
-contact between its associated geom and other geoms. The vertices of the mesh, when positioned in the geom frame, are
-the points at which sensor values are computed, so the dimension of the output is 3 times the number of vertices in the
-mesh. The mesh must have 3 normal vectors per vertex, which are used to compute the tangent frame. If the penetration
-depth is positive (no contact), then all values are 0 for the corresponding vertex. Only contacts with geoms of type SDF
-contribute to the sensor output. The sensor can be visualized by enabling the visualization of contact points.
+The tactile sensor returns the maximum penetration depth and the sliding velocities in the tangent frame at given points
+between the geom associated with the sensor and the SDF geoms in contact with it. The sensor is associated with a geom
+and a mesh. It is activated by the contact between its associated geom and other geoms. The vertices of the mesh, when
+positioned in the geom frame, are the points at which sensor values are computed, so the dimension of the output is 3
+times the number of vertices in the mesh. The mesh must have 3 normal vectors per vertex, which are used to compute the
+tangent frame. If the penetration depth is positive (no contact), then all values are 0 for the corresponding vertex.
+Only contacts with geoms of type SDF contribute to the sensor output. The sensor can be visualized by enabling the
+visualization of contact points.
 
 .. _sensor-tactile-geom:
 

doc/changelog.rst

@@ -79,6 +79,8 @@ General
      `mjtype.h <https://github.com/google-deepmind/mujoco/blob/main/include/mujoco/mjtype.h>` and now includes all
      enum type definitions.
 
+   - The :ref:`tactile<sensor-tactile>` sensor now reports raw depth instead of an estimated pressure.
+
    - MJX: Removed the deprecated ``nconmax`` argument from ``mjx.make_data`` and ``mjx.put_data`` in favor of
      ``naconmax``.
 

mjx/mujoco/mjx/third_party/mujoco_warp/_src/sensor.py

@@ -2252,18 +2252,14 @@ def _sensor_tactile(
     )
     vel_rel = vel_sensor - vel_other
 
-    kMaxDepth = 0.05
-    pressure = depth / wp.max(kMaxDepth - depth, MJ_MINVAL)
-    force = wp.mul(normal, pressure)
-
     forceT = wp.vec3(0.0, 0.0, 0.0)
-    forceT[0] = wp.dot(force, normal)
+    forceT[0] = -depth
     if has_frame:
       forceT[1] = wp.abs(wp.dot(vel_rel, tang1))
       forceT[2] = wp.abs(wp.dot(vel_rel, tang2))
 
     dim = sensor_dim[sensor_id] // 3
-    wp.atomic_add(sensordata_out, worldid, sensor_adr[sensor_id] + 0 * dim + vertid, forceT[0])
+    wp.atomic_max(sensordata_out, worldid, sensor_adr[sensor_id] + 0 * dim + vertid, forceT[0])
     wp.atomic_add(sensordata_out, worldid, sensor_adr[sensor_id] + 1 * dim + vertid, forceT[1])
     wp.atomic_add(sensordata_out, worldid, sensor_adr[sensor_id] + 2 * dim + vertid, forceT[2])
 

src/engine/engine_sensor.c

@@ -172,14 +172,8 @@ static void* tactile_taxel_batch(void* args) {
                           mesh_normal[normal_stride*j + 2]};
       mju_rotVecQuat(normal, normal, m->mesh_quat + 4 * mesh_id);
 
-      // get contact force
-      mjtNum force[3];
-      mjtNum kMaxDepth = 0.05;
-      mjtNum pressure = depth / mju_max(kMaxDepth - depth, mjMINVAL);
-      mju_scl3(force, normal, pressure);
-
-      // accumulate into forcesT (disjoint writes per taxel j)
-      t->forcesT[0*ncon + j] += mju_dot3(force, normal);
+      // take max penetration depth (SDF distance is negative; negate for positive output)
+      t->forcesT[0*ncon + j] = mju_max(t->forcesT[0*ncon + j], -depth);
 
       if (has_frame) {
         mjtNum tang1[3] = {mesh_normal[normal_stride*j + 3],

test/engine/engine_sensor_test.cc

@@ -1702,7 +1702,7 @@ TEST_F(SensorTest, TactileSkipTangents) {
   EXPECT_EQ(data->time, 0.0);
   EXPECT_GT(data->ncon, 0) << "No contacts generated";
 
-  // Tactile sensor layout: [normal_forces..., tang1_forces..., tang2_forces...]
+  // Tactile sensor layout: [depths..., tang1_vel..., tang2_vel...]
   int ntaxel = model->nsensordata / 3;
   ASSERT_EQ(model->nsensordata % 3, 0) << "Sensor dim should be divisible by 3";
 
@@ -1712,13 +1712,13 @@ TEST_F(SensorTest, TactileSkipTangents) {
         << "Tangent component at index " << i << " should be 0";
   }
 
-  // Normal force components: verify count, sign, and magnitude ~-0.8
+  // Penetration depth components: verify count, sign, and magnitude ~0.2
   int nonzero_count = 0;
   for (int i = 0; i < ntaxel; i++) {
     if (data->sensordata[i] != 0) {
       nonzero_count++;
-      EXPECT_NEAR(data->sensordata[i], -0.8, 0.1)
-          << "Normal force at taxel " << i;
+      EXPECT_NEAR(data->sensordata[i], 0.2, 0.1)
+          << "Penetration depth at taxel " << i;
     }
   }
   EXPECT_EQ(nonzero_count, 2) << "Expected 2 taxels in contact";