Re-enable add_base_mass on H1/Cassie with log-uniform scale

Restore base-mass randomization on H1 and Cassie that was disabled with
add_base_mass = None in their rough-terrain configs (pre-existing biped
convention from 2024-06 PR #444, later reinforced by PR #4165's Newton NaN
TODO). The parent PR #5248 switches the shared default to log-uniform scale
(1/1.25, 1.25), which is safer for bipeds than the old additive (-5, 5) kg
(effectively ±25% on H1's torso vs ±100% on Cassie's pelvis).

- H1 inherits the shared default (symmetric ±25% scale, body_names="torso_link").
- Cassie overrides to (1.0, 1.25) asymmetric heavier-bias: lighter-than-
  nominal pelvis destabilizes Cassie's closed-loop Achilles rod coupling
  and hip PD response, while heavier-than-nominal dampens dynamics.

| Variant                        | reward | ep len | vs disabled |
|--------------------------------|-------:|-------:|------------:|
| None (disabled, prior default) |  23.93 |    910 |     1.00x   |
| (1/1.25, 1.25) sym25           |  14.15 |    850 |     0.59x x |
| (1/1.10, 1.10) tight10         |  14.53 |    831 |     0.61x x |
| **(1.0, 1.25) heavier25**      | **21.50** | **932** | **0.90x** |

Tightening the range symmetrically did not help (tight10 ~ sym25) — the
lighter side is what destabilizes, not the magnitude. Restricting to
[1.0, 1.25] (never lighter, up to +25% heavier) preserves most of the
randomization benefit while avoiding the failure mode. Episode length
also improves (932 vs 910), indicating the randomized policy is actively
more stable during episodes — the lower aggregate reward comes from extra
dof-torque regularizer paid to handle heavier instances, not degraded
task completion.

H1's larger base mass (≈15 kg torso) means ±25% on the default scale
(11-19 kg range) is well within the controller's robustness margin. H1
reward at iter 1499: 24.02 with mass rand on vs 23.58 with it disabled
(1.02x, essentially equal). Re-enabling provides sim-to-real robustness
at negligible training cost.

Author: hujc7

source/isaaclab_tasks/config/extension.toml

@@ -1,7 +1,7 @@
 [package]
 
 # Note: Semantic Versioning is used: https://semver.org/
-version = "1.5.25"
+version = "1.5.26"
 
 # Description
 title = "Isaac Lab Environments"

source/isaaclab_tasks/docs/CHANGELOG.rst

@@ -1,6 +1,23 @@
 Changelog
 ---------
 
+1.5.26 (2026-04-24)
+~~~~~~~~~~~~~~~~~~~
+
+Changed
+^^^^^^^
+
+* Re-enabled ``add_base_mass`` randomization on H1 and Cassie in their
+  rough-terrain configs (previously ``= None`` per the pre-existing biped
+  convention). H1 uses the shared log-uniform scale default from
+  ``EventsCfg``; Cassie overrides to ``(1.0, 1.25)`` asymmetric heavier-bias
+  (never lighter than nominal). Symmetric ±25% regressed Cassie reward by
+  40% vs disabled due to closed-loop Achilles coupling destabilizing on
+  lighter pelvis mass; ``(1.0, 1.25)`` recovers to 90% of the
+  mass-rand-disabled baseline while retaining the domain-randomization
+  benefit.
+
+
 1.5.25 (2026-04-24)
 ~~~~~~~~~~~~~~~~~~~
 

source/isaaclab_tasks/isaaclab_tasks/manager_based/locomotion/velocity/config/cassie/rough_env_cfg.py

@@ -67,8 +67,14 @@ def __post_init__(self):
 
         self.scene.height_scanner.prim_path = "{ENV_REGEX_NS}/Robot/pelvis"
 
-        # Cassie uses "pelvis" as base body — disable mass randomization for bipeds
-        self.events.add_base_mass = None
+        # Cassie uses "pelvis" as base body. Override the shared symmetric
+        # (1/1.25, 1.25) log-uniform scale with asymmetric (1.0, 1.25) —
+        # lighter-than-nominal pelvis destabilizes Cassie's closed-loop
+        # Achilles coupling + hip PD response, so only heavier perturbations
+        # are safe. Symmetric ±25% regressed reward 40% vs disabled;
+        # (1.0, 1.25) recovers to 90% of baseline.
+        self.events.add_base_mass.params["asset_cfg"].body_names = "pelvis"
+        self.events.add_base_mass.params["mass_distribution_params"] = (1.0, 1.25)
         self.events.base_com = None
         self.events.base_external_force_torque.params["asset_cfg"].body_names = ".*pelvis"
         # Cassie has precise initial pose — don't scale joint defaults randomly on reset

source/isaaclab_tasks/isaaclab_tasks/manager_based/locomotion/velocity/config/h1/rough_env_cfg.py

@@ -83,8 +83,9 @@ def __post_init__(self):
         if self.scene.height_scanner:
             self.scene.height_scanner.prim_path = "{ENV_REGEX_NS}/Robot/torso_link"
 
-        # H1 uses "torso_link" as base body — disable mass randomization for bipeds
-        self.events.add_base_mass = None
+        # H1 uses "torso_link" as base body; inherits the shared log-uniform mass
+        # randomization scale from EventsCfg (no per-H1 override needed).
+        self.events.add_base_mass.params["asset_cfg"].body_names = "torso_link"
         # H1 has precise initial pose — don't scale joint defaults randomly on reset
         self.events.reset_robot_joints.params["position_range"] = (1.0, 1.0)
         self.events.base_com = None