Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Randomization, noise and delay profiles

SRB configures three things separately: event randomization, action and observation delay, and Isaac Lab action-noise models.

Task/configRandomizationDelayNoiseProvenance
waypoint_navigationtarget_pose_evolution, a tagged task event at a 0.05 s intervalAction (0, 3) and observation (0, 1) agent steps; per-reset draws use env.seedWaypoint-relative observation noise: 0.01 m / 2.5° episode bias plus 0.0025 m / 0.5° per-step noiseTask defaults; uncalibrated
locomotion_waypoint_navigationInherits target_pose_evolutionInherits action (0, 3) and observation (0, 1) agent stepsInherits waypoint-relative observation noise: 0.01 m / 2.5° episode bias plus 0.0025 m / 0.5° per-step noiseInherited waypoint task config; uncalibrated
waypoint_navigation with env.sim_to_real_profile=waypoint_navigation_synthetic_v1Same target eventAction and observation (1, 2) agent steps, fixed for each episode; 25 Hz means 40–80 msGaussian action noise, standard deviation 0.02 in normalized policy-action units; the waypoint observation noise above remains in metres and radianssynthetic_unmeasured
Other shipped task configsMany define tagged or untagged event termsNo other task config enables nonzero action or observation delayNo default task config enables Isaac Lab action_noise_model or observation_noise_modelPer-task terms

Opt in to the synthetic profile:

srb agent train --algo sbx_ppo --env waypoint_navigation \
  env.sim_to_real_profile=waypoint_navigation_synthetic_v1 \
  env.seed=20260908

The base config sets env.seed=0. default_seed applies only when the task config has env.seed=None; any explicit seed, including 0, wins.

Delay ranges count integer agent steps. At 25 Hz, one step is 0.04 s. On reset, each environment samples its delay. Observation-delay history starts with the fresh observation; action-delay history starts with zeros.

Override delay and noise fields only with values matching the task default or the profile; other values fail. Zero change probability disables boundary redraws, so subset resets cannot reuse another environment’s history. The profile requires agent_rate=0.04 seconds and rejects custom action_noise_model to keep delay and noise units fixed.

The synthetic_unmeasured profile has synthetic delays and Gaussian action noise, not hardware calibration, sensor statistics, or latency measurements.

RandomizationView exposes active event terms and reports untagged EventTermCfg fields as UNTAGGED. The profile does not retag other task configs or enable noise globally.