diff --git a/examples/MultiThreading/MultiThreadingExample.cpp b/examples/MultiThreading/MultiThreadingExample.cpp
index c63a0b363..b9701fb7e 100644
--- a/examples/MultiThreading/MultiThreadingExample.cpp
+++ b/examples/MultiThreading/MultiThreadingExample.cpp
@@ -29,13 +29,14 @@ void SamplelsMemoryReleaseFunc(void* ptr);
 #ifndef _WIN32
 #include "b3PosixThreadSupport.h"
 
+
 b3ThreadSupportInterface* createThreadSupport(int numThreads)
 {
 	b3PosixThreadSupport::ThreadConstructionInfo constructionInfo("testThreads",
-																  SampleThreadFunc,
-																  SamplelsMemoryFunc,
-																  SamplelsMemoryReleaseFunc,
-																  numThreads);
+	SampleThreadFunc,
+	SamplelsMemoryFunc,
+	SamplelsMemoryReleaseFunc,
+	  numThreads);
 	b3ThreadSupportInterface* threadSupport = new b3PosixThreadSupport(constructionInfo);
 
 	return threadSupport;
diff --git a/examples/MultiThreading/main.cpp b/examples/MultiThreading/main.cpp
index f2dcd13dc..561f17c21 100644
--- a/examples/MultiThreading/main.cpp
+++ b/examples/MultiThreading/main.cpp
@@ -22,7 +22,7 @@ subject to the following restrictions:
 
 void SampleThreadFunc(void* userPtr, void* lsMemory);
 void* SamplelsMemoryFunc();
-
+void SamplelsMemoryReleaseFunc(void* ptr);
 #include <stdio.h>
 //#include "BulletMultiThreaded/PlatformDefinitions.h"
 
@@ -34,6 +34,7 @@ b3ThreadSupportInterface* createThreadSupport(int numThreads)
 	b3PosixThreadSupport::ThreadConstructionInfo constructionInfo("testThreads",
 																  SampleThreadFunc,
 																  SamplelsMemoryFunc,
+SamplelsMemoryReleaseFunc,
 																  numThreads);
 	b3ThreadSupportInterface* threadSupport = new b3PosixThreadSupport(constructionInfo);
 
@@ -99,6 +100,13 @@ void* SamplelsMemoryFunc()
 	return new SampleThreadLocalStorage;
 }
 
+void SamplelsMemoryReleaseFunc(void* ptr)
+{
+        SampleThreadLocalStorage* p = (SampleThreadLocalStorage*)ptr;
+        delete p;
+}
+
+
 int main(int argc, char** argv)
 {
 	int numThreads = 8;
diff --git a/examples/SharedMemory/PhysicsClientC_API.cpp b/examples/SharedMemory/PhysicsClientC_API.cpp
index ee506c143..e6c5aeecc 100644
--- a/examples/SharedMemory/PhysicsClientC_API.cpp
+++ b/examples/SharedMemory/PhysicsClientC_API.cpp
@@ -404,7 +404,7 @@ B3_SHARED_API int b3LoadSoftBodySetCollisionHardness(b3SharedMemoryCommandHandle
 	return 0;
 }
 
-B3_SHARED_API int b3LoadSoftBodyUseSelfCollision(b3SharedMemoryCommandHandle commandHandle, int useSelfCollision)
+B3_SHARED_API int b3LoadSoftBodySetSelfCollision(b3SharedMemoryCommandHandle commandHandle, int useSelfCollision)
 {
 	struct SharedMemoryCommand* command = (struct SharedMemoryCommand*)commandHandle;
 	b3Assert(command->m_type == CMD_LOAD_SOFT_BODY);
diff --git a/examples/SharedMemory/PhysicsClientC_API.h b/examples/SharedMemory/PhysicsClientC_API.h
index c67f79d9c..856fa2cc4 100644
--- a/examples/SharedMemory/PhysicsClientC_API.h
+++ b/examples/SharedMemory/PhysicsClientC_API.h
@@ -640,7 +640,7 @@ extern "C"
 	B3_SHARED_API int b3LoadSoftBodyAddMassSpringForce(b3SharedMemoryCommandHandle commandHandle, double springElasticStiffness , double springDampingStiffness);
 	B3_SHARED_API int b3LoadSoftBodyAddGravityForce(b3SharedMemoryCommandHandle commandHandle, double gravityX, double gravityY, double gravityZ);
 	B3_SHARED_API int b3LoadSoftBodySetCollisionHardness(b3SharedMemoryCommandHandle commandHandle, double collisionHardness);
-	B3_SHARED_API int b3LoadSoftBodyUseSelfCollision(b3SharedMemoryCommandHandle commandHandle, int useSelfCollision);
+	B3_SHARED_API int b3LoadSoftBodySetSelfCollision(b3SharedMemoryCommandHandle commandHandle, int useSelfCollision);
     B3_SHARED_API int b3LoadSoftBodyUseFaceContact(b3SharedMemoryCommandHandle commandHandle, int useFaceContact);
 	B3_SHARED_API int b3LoadSoftBodySetFrictionCoefficient(b3SharedMemoryCommandHandle commandHandle, double frictionCoefficient);
 	B3_SHARED_API int b3LoadSoftBodyUseBendingSprings(b3SharedMemoryCommandHandle commandHandle, int useBendingSprings);
diff --git a/examples/pybullet/gym/pybullet_envs/deep_mimic/env/humanoid_stable_pd.py b/examples/pybullet/gym/pybullet_envs/deep_mimic/env/humanoid_stable_pd.py
index cbdf9d191..5a07d70a6 100644
--- a/examples/pybullet/gym/pybullet_envs/deep_mimic/env/humanoid_stable_pd.py
+++ b/examples/pybullet/gym/pybullet_envs/deep_mimic/env/humanoid_stable_pd.py
@@ -807,7 +807,10 @@ class HumanoidStablePD(object):
     rootPosSim, rootOrnSim = self._pybullet_client.getBasePositionAndOrientation(self._sim_model)
     rootPosKin, rootOrnKin = self._pybullet_client.getBasePositionAndOrientation(self._kin_model)
     linVelSim, angVelSim = self._pybullet_client.getBaseVelocity(self._sim_model)
-    linVelKin, angVelKin = self._pybullet_client.getBaseVelocity(self._kin_model)
+    #don't read the velocities from the kinematic model (they are zero), use the pose interpolator velocity
+    #see also issue https://github.com/bulletphysics/bullet3/issues/2401 
+    linVelKin = self._poseInterpolator._baseLinVel
+    angVelKin = self._poseInterpolator._baseAngVel
 
     root_rot_err = self.calcRootRotDiff(rootOrnSim, rootOrnKin)
     pose_err += root_rot_w * root_rot_err
diff --git a/setup.py b/setup.py
index b84bc9385..ae78e4690 100644
--- a/setup.py
+++ b/setup.py
@@ -491,7 +491,7 @@ if 'BT_USE_EGL' in EGL_CXX_FLAGS:
 
 setup(
     name='pybullet',
-    version='2.5.7',
+    version='2.5.8',
     description=
     'Official Python Interface for the Bullet Physics SDK specialized for Robotics Simulation and Reinforcement Learning',
     long_description=