diff --git a/python/ray/air/_internal/tensorflow_utils.py b/python/ray/air/_internal/tensorflow_utils.py
index 13e947dc5..d81e600ab 100644
--- a/python/ray/air/_internal/tensorflow_utils.py
+++ b/python/ray/air/_internal/tensorflow_utils.py
@@ -47,6 +47,13 @@ def convert_pandas_to_tf_tensor(
             # them. If the columns contain different types (for example, `float32`s
             # and `int32`s), then `tf.concat` raises an error.
             dtype: np.dtype = np.find_common_type(df.dtypes, [])
+
+            # if the columns are `ray.data.extensions.tensor_extension.TensorArray`,
+            # the dtype will be `object`. In this case, we need to set the dtype to
+            # none, and use the automatic type casting of `tf.convert_to_tensor`.
+            if isinstance(dtype, object):
+                dtype = None
+
         except TypeError:
             # `find_common_type` fails if a series has `TensorDtype`. In this case,
             # don't cast any of the series and continue.