Bump G1 max_iterations to 6000 on Newton for rough-terrain parity

Newton converges ~2x slower than PhysX on G1 rough terrain. Both
backends reach the same reward plateau (r~+17-19): PhysX near iter
3000, Newton near iter 6000. Ablation (armature 0.01/0.03, damping
tuning, Newton a27277 upstream) did not close the gap — the cost is
sample efficiency, not a ceiling.

Use the framework preset on max_iterations rather than tuning physics
or reward terms, keeping the env config engine-agnostic. Precedent:
Allegro Hand (5000), Spot (20000).

Author: hujc7

source/isaaclab_tasks/config/extension.toml

@@ -1,7 +1,7 @@
 [package]
 
 # Note: Semantic Versioning is used: https://semver.org/
-version = "1.5.23"
+version = "1.5.24"
 
 # Description
 title = "Isaac Lab Environments"

source/isaaclab_tasks/docs/CHANGELOG.rst

@@ -1,6 +1,19 @@
 Changelog
 ---------
 
+1.5.24 (2026-04-22)
+~~~~~~~~~~~~~~~~~~~
+
+Added
+^^^^^
+
+* Added Newton rough terrain support for the G1 biped locomotion velocity
+  env. The only engine-specific change is a 2x ``max_iterations`` preset on
+  :class:`~isaaclab_tasks.manager_based.locomotion.velocity.config.g1.agents.rsl_rl_ppo_cfg.G1RoughPPORunnerCfg`
+  (Newton = 6000, PhysX = 3000). Both backends converge to the same reward
+  plateau (r≈+17-19); Newton is phase-shifted ~2000 iter later, so the
+  iteration budget is bumped rather than tuning physics or reward terms.
+
 1.5.23 (2026-04-22)
 ~~~~~~~~~~~~~~~~~~~
 

source/isaaclab_tasks/isaaclab_tasks/manager_based/locomotion/velocity/config/g1/agents/rsl_rl_ppo_cfg.py

@@ -6,12 +6,16 @@
 from isaaclab.utils import configclass
 
 from isaaclab_rl.rsl_rl import RslRlOnPolicyRunnerCfg, RslRlPpoActorCriticCfg, RslRlPpoAlgorithmCfg
+from isaaclab_tasks.utils import preset
 
 
 @configclass
 class G1RoughPPORunnerCfg(RslRlOnPolicyRunnerCfg):
     num_steps_per_env = 24
-    max_iterations = 3000
+    # Newton needs ~2x the PPO iterations to reach the same reward plateau as PhysX on G1.
+    # Both backends converge to r≈+17-19; PhysX hits it near iter 3000, Newton near iter 6000.
+    # The gap is sample-efficiency, not a ceiling — no physics or reward tuning closes it.
+    max_iterations = preset(default=3000, newton=6000)
     save_interval = 50
     experiment_name = "g1_rough"
     policy = RslRlPpoActorCriticCfg(