diff --git a/NEWS.md b/NEWS.md
index 6b6999c7a6..0a127fae71 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -232,6 +232,8 @@
 
 29. `integer64` defined on a subset of a new column would leave "gibberish" on the remaining rows, [#3723](https://github.com/Rdatatable/data.table/issues/3723). A bug in `rbindlist` with the same root cause was also fixed, [#1459](https://github.com/Rdatatable/data.table/issues/1459). Thanks @shrektan and @jangorecki for the reports.
 
+30. `groupingsets` functions now properly handle alone special symbols when using an empty set to group by, [#3653](https://github.com/Rdatatable/data.table/issues/3653). Thanks to @Henrik-P for the report.
+
 #### NOTES
 
 1. `rbindlist`'s `use.names="check"` now emits its message for automatic column names (`"V[0-9]+"`) too, [#3484](https://github.com/Rdatatable/data.table/pull/3484). See news item 5 of v1.12.2 below.
diff --git a/R/groupingsets.R b/R/groupingsets.R
index a13d71cd50..dadcbb0fe8 100644
--- a/R/groupingsets.R
+++ b/R/groupingsets.R
@@ -72,15 +72,7 @@ groupingsets.data.table = function(x, j, by, sets, .SDcols, id = FALSE, jj, ...)
   if (missing(.SDcols))
     .SDcols = if (".SD" %chin% av) setdiff(names(x), by) else NULL
   # 0 rows template data.table to keep colorder and type
-  if (length(by)) {
-    empty = if (length(.SDcols)) x[0L, eval(jj), by, .SDcols=.SDcols] else x[0L, eval(jj), by]
-  } else {
-    empty = if (length(.SDcols)) x[0L, eval(jj), .SDcols=.SDcols] else x[0L, eval(jj)]
-    if (!is.data.table(empty)) {
-      if (length(empty)>0) empty = empty[0L] # fix for #3173 when no grouping and j constant
-      empty = setDT(list(empty)) # improve after #648, see comment in aggregate.set
-    }
-  }
+  empty = if (length(.SDcols)) x[0L, eval(jj), by, .SDcols=.SDcols] else x[0L, eval(jj), by]
   if (id && "grouping" %chin% names(empty)) # `j` could have been evaluated to `grouping` field
     stop("When using `id=TRUE` the 'j' expression must not evaluate to a column named 'grouping'.")
   if (anyDuplicated(names(empty)) > 0L)
@@ -98,13 +90,7 @@ groupingsets.data.table = function(x, j, by, sets, .SDcols, id = FALSE, jj, ...)
   int64.by.cols = intersect(int64.cols, by)
   # aggregate function called for each grouping set
   aggregate.set = function(by.set) {
-    if (length(by.set)) {
-      r = if (length(.SDcols)) x[, eval(jj), by.set, .SDcols=.SDcols] else x[, eval(jj), by.set]
-    } else {
-      r = if (length(.SDcols)) x[, eval(jj), .SDcols=.SDcols] else x[, eval(jj)]
-      # workaround for grand total single var as data.table too, change to drop=FALSE after #648 solved
-      if (!is.data.table(r)) r = setDT(list(r))
-    }
+    r = if (length(.SDcols)) x[, eval(jj), by.set, .SDcols=.SDcols] else x[, eval(jj), by.set]
     if (id) {
       # integer bit mask of aggregation levels: http://www.postgresql.org/docs/9.5/static/functions-aggregate.html#FUNCTIONS-GROUPING-TABLE
       # 3267: strtoi("", base = 2L) output apparently unstable across platforms
diff --git a/inst/tests/tests.Rraw b/inst/tests/tests.Rraw
index aade8300e6..af01a1addd 100644
--- a/inst/tests/tests.Rraw
+++ b/inst/tests/tests.Rraw
@@ -15738,6 +15738,15 @@ if (test_bit64) {
   test(2083.2, rbind(data.table(a=1:2, b=as.integer64(c(1,NA))), data.table(a=3L), fill=TRUE)$b, as.integer64(c(1, NA, NA)))
 }
 
+# groupingsets j=.N by character(0) set #3653
+d = data.table(x = c("a", "a", "b"))
+test(2084.01, groupingsets(d, j = .N, by = "x", sets = list("x", character())), data.table(x=c("a","b",NA_character_), N=c(2L,1L,3L)))
+test(2084.02, groupingsets(d, j = .N, by = "x", sets = list(character())), data.table(x=NA_character_, N=3L))
+test(2084.03, groupingsets(d, j = .GRP, by = "x", sets = list("x", character())), data.table(x=c("a","b",NA_character_), GRP=c(1L,2L,1L)))
+test(2084.04, groupingsets(d, j = .GRP, by = "x", sets = list(character())), data.table(x=NA_character_, GRP=1L))
+test(2084.05, groupingsets(d, j = .I, by = "x", sets = list("x", character())), data.table(x=c("a","a","b",rep(NA_character_,3L)), I=c(1:3,1:3)))
+test(2084.06, groupingsets(d, j = .I, by = "x", sets = list(character())), data.table(x=rep(NA_character_,3L), I=1:3))
+
 
 ###################################
 #  Add new tests above this line  #