[Bug Report] Slow offline rendering during training

[jnskkmhr]

Questions

1. Is continuous rendering with --video the intended behavior, or is there an existing way to restrict
   rendering (and camera pipeline updates) to the active recording windows only?
2. Is there a recommended alternative for periodic offline visual monitoring under the Newton backend that
   avoids the RTX pipeline entirely (e.g. periodic Newton viewer / Rerun captures)?

Environment

• Isaac Lab: develop (Newton physics backend, MJWarp solver)
• Isaac Sim: 6.0.x
• Setup: headless training in Docker, RSL-RL
• Task: manager-based humanoid locomotion, dt=0.005, decimation=4, render_interval=4
• Repro: train.py --task <task> --headless --viz none --video --video_length 200 --video_interval 2000
  → ~10 s/iteration vs ~1 s/iteration without --video

[RandomOakForest]

Thank you for posting this. The team is addressing the issue.