3.0.0b2: ~20% Humanoid-Direct throughput regression vs v2.3.2 stack (Shadow-Cube unaffected)

Migrating from the Isaac Lab v2.3.2 stack to 3.0.0b2 costs ~20% of Humanoid-Direct training throughput at scale, while Shadow-Cube gets slightly faster. All numbers below are measured with the two stacks equalized down to the RL code: same torch 2.11.0 with identical nvidia-* libs on both sides, the same rl_games code forced onto both, identical task configs (dt, decimation, PhysX solver and GPU buffer settings verified field by field; zero PhysX buffer overflows in any run), exclusive machine, back-to-back runs.

Setup: RTX PRO 6000 Blackwell Max-Q (driver 595.84), Threadripper 9970X, Ubuntu 24.04. Old = Isaac Lab v2.3.2 + Isaac Sim 5.1.0. New = isaaclab 3.0.0b2.post1 + Isaac Sim 6.0.1. rl_games PPO, 300-iteration runs, single GPU.

RL collection throughput (step FPS, rollout phase only, no optimizer time):

task	envs	2.3.2	3.0.0b2	delta
Humanoid-Direct	16384	750k	599k	-20%
Humanoid-Direct	8192	622k	551k	-11%
Humanoid-Direct	4096	516k	445k	-14%
Shadow-Cube (Repose)	16384	311k	319k	+2%
Shadow-Cube (Repose)	8192	249k	257k	+3%
Shadow-Cube (Repose)	4096	183k	193k	+6%

Pure sim without RL (stepping DirectRLEnv with random actions, 1000 timed steps): Humanoid 1267k -> 1185k (-6%) at 16K, 971k -> 925k (-5%) at 8K; Shadow 528k -> 517k (-2%), 394k -> 384k (-3%).

Where the loss is: the sim itself is only 2-6% slower, and Shadow actually trains slightly faster on 3.0 — but Humanoid loses ~20% at scale. The loss sits outside env.step(): per-env-step time minus pure-sim time is ~0.53us on 2.3.2 vs ~0.77us on 3.0.0b2, roughly constant across env counts — a flat per-env tax in the collection path whose relative cost grows with env count. It is not policy inference and not the RL library (both bit-identical across the comparison). GPU utilization drops accordingly (39% -> 33% at 16K for the same workload).

Looks like per-env overhead in the 3.0 env/wrapper plumbing that locomotion-style envs are most exposed to. Can profile the collection loop or test patches — the hardware stays available.