Bug: Intel XPU training fails with "UR error" in rsl_rl

[wpjNB]

Summary

When training with Intel XPU device (training.device=xpu), the training fails with
RuntimeError: UR error from Intel's Unified Runtime. The error occurs in rsl_rl's
normalization and storage modules when performing division and .copy_() operations on
XPU tensors.

Area

algos

Reproduction

Install XPU torch

make sync-xpu

Run training

uv run --no-sync train --algo ppo --task go2_joystick_flat --sim mujoco training.device=xpu training.no_play=true

Expected behavior

Training should work with Intel XPU device without runtime errors.

Actual behavior

Training fails with RuntimeError: UR error during the first training iteration. The
error occurs in two places:

1. rsl_rl/modules/normalization.py:61 - Division operation in update() method
2. rsl_rl/storage/rollout_storage.py:180 - .copy_() operation in add_transition()
   method

Environment

• OS: WSL2 (Ubuntu 22.04)
• Python: 3.10
• Backend: MuJoCo
• Device: Intel XPU (detected via OpenCL)
• PyTorch: 2.7.0+xpu
• rsl_rl: 5.0.1
• Commit: 70d8f48

Logs and artifacts