Expands observation term scaling to support list of floats

source/extensions/omni.isaac.lab/config/extension.toml

@@ -2,7 +2,7 @@
 
 # Note: Semantic Versioning is used: https://semver.org/
 
-version = "0.27.7"
+version = "0.27.8"
 
 # Description
 title = "Isaac Lab framework for Robot Learning"

source/extensions/omni.isaac.lab/docs/CHANGELOG.rst

@@ -1,6 +1,15 @@
 Changelog
 ---------
 
+0.27.8 (2024-10-31)
+~~~~~~~~~~~~~~~~~~~~
+
+Added
+^^^^^
+
+* Added support to define tuple of floats to scale observation terms by expanding the
+  :attr:`omni.isaac.lab.managers.manager_term_cfg.ObservationManagerCfg.scale` attribute.
+
 
 0.27.7 (2024-10-28)
 ~~~~~~~~~~~~~~~~~~~

source/extensions/omni.isaac.lab/omni/isaac/lab/managers/manager_term_cfg.py

@@ -152,9 +152,13 @@ class ObservationTermCfg(ManagerTermBaseCfg):
     """The clipping range for the observation after adding noise. Defaults to None,
     in which case no clipping is applied."""
 
-    scale: float | None = None
+    scale: tuple[float, ...] | float | None = None
     """The scale to apply to the observation after clipping. Defaults to None,
-    in which case no scaling is applied (same as setting scale to :obj:`1`)."""
+    in which case no scaling is applied (same as setting scale to :obj:`1`).
+
+    We leverage PyTorch broadcasting to scale the observation tensor with the provided value. If a tuple is provided,
+    please make sure the length of the tuple matches the dimensions of the tensor outputted from the term.
+    """
 
 
 @configclass

source/extensions/omni.isaac.lab/omni/isaac/lab/managers/observation_manager.py

@@ -259,7 +259,7 @@ def compute_group(self, group_name: str) -> torch.Tensor | dict[str, torch.Tenso
                 obs = term_cfg.noise.func(obs, term_cfg.noise)
             if term_cfg.clip:
                 obs = obs.clip_(min=term_cfg.clip[0], max=term_cfg.clip[1])
-            if term_cfg.scale:
+            if term_cfg.scale is not None:
                 obs = obs.mul_(term_cfg.scale)
             # add value to list
             group_obs[name] = obs
@@ -343,6 +343,23 @@ def _prepare_terms(self):
                 obs_dims = tuple(term_cfg.func(self._env, **term_cfg.params).shape)
                 self._group_obs_term_dim[group_name].append(obs_dims[1:])
 
+                # if scale is set, check if single float or tuple
+                if term_cfg.scale is not None:
+                    if not isinstance(term_cfg.scale, (float, int, tuple)):
+                        raise TypeError(
+                            f"Scale for observation term '{term_name}' in group '{group_name}'"
+                            f" is not of type float, int or tuple. Received: '{type(term_cfg.scale)}'."
+                        )
+                    if isinstance(term_cfg.scale, tuple) and len(term_cfg.scale) != obs_dims[1]:
+                        raise ValueError(
+                            f"Scale for observation term '{term_name}' in group '{group_name}'"
+                            f" does not match the dimensions of the observation. Expected: {obs_dims[1]}"
+                            f" but received: {len(term_cfg.scale)}."
+                        )
+
+                    # cast the scale into torch tensor
+                    term_cfg.scale = torch.tensor(term_cfg.scale, dtype=torch.float, device=self._env.device)
+
                 # prepare modifiers for each observation
                 if term_cfg.modifiers is not None:
                     # initialize list of modifiers for term

source/extensions/omni.isaac.lab/test/managers/test_observation_manager.py

@@ -244,6 +244,8 @@ class SampleImageGroupCfg(ObservationGroupCfg):
     def test_compute(self):
         """Test the observation computation."""
 
+        pos_scale_tuple = (2.0, 3.0, 1.0)
+
         @configclass
         class MyObservationManagerCfg:
             """Test config class for observation manager."""
@@ -254,14 +256,14 @@ class PolicyCfg(ObservationGroupCfg):
 
                 term_1 = ObservationTermCfg(func=grilled_chicken, scale=10)
                 term_2 = ObservationTermCfg(func=grilled_chicken_with_curry, scale=0.0, params={"hot": False})
-                term_3 = ObservationTermCfg(func=pos_w_data, scale=2.0)
+                term_3 = ObservationTermCfg(func=pos_w_data, scale=pos_scale_tuple)
                 term_4 = ObservationTermCfg(func=lin_vel_w_data, scale=1.5)
 
             @configclass
             class CriticCfg(ObservationGroupCfg):
-                term_1 = ObservationTermCfg(func=pos_w_data, scale=2.0)
+                term_1 = ObservationTermCfg(func=pos_w_data, scale=pos_scale_tuple)
                 term_2 = ObservationTermCfg(func=lin_vel_w_data, scale=1.5)
-                term_3 = ObservationTermCfg(func=pos_w_data, scale=2.0)
+                term_3 = ObservationTermCfg(func=pos_w_data, scale=pos_scale_tuple)
                 term_4 = ObservationTermCfg(func=lin_vel_w_data, scale=1.5)
 
             @configclass
@@ -289,6 +291,11 @@ class ImageCfg(ObservationGroupCfg):
         self.assertEqual((self.env.num_envs, 11), obs_policy.shape)
         self.assertEqual((self.env.num_envs, 12), obs_critic.shape)
         self.assertEqual((self.env.num_envs, 128, 256, 4), obs_image.shape)
+        # check that the scales are applied correctly
+        torch.testing.assert_close(
+            self.env.data.pos_w * torch.tensor(pos_scale_tuple, device=self.env.device), obs_critic[:, :3]
+        )
+        torch.testing.assert_close(self.env.data.lin_vel_w * 1.5, obs_critic[:, 3:6])
         # make sure that the data are the same for same terms
         # -- within group
         torch.testing.assert_close(obs_critic[:, 0:3], obs_critic[:, 6:9])