Sim2Real Transfer

Tools for narrowing the gap between simulated training and real-vehicle deployment.

Domain Randomization

At each reset, selected vehicle parameters are perturbed by multiplicative Gaussian noise:

\[\theta' = \theta \cdot X, \quad X \sim \mathcal{N}(1, \sigma)\]

clipped at \(\pm\) dr_clip_k \(\cdot \sigma\) (default dr_clip_k = 3.0). This widens the distribution of dynamics the policy sees during training, improving robustness to unknown real-world parameter offsets.

Configured via the domain_randomization gym key as a flat {param: sigma} dict (\(\sigma\) in \((0, 0.2)\)); absent or None disables DR. Train configs (get_drift_train_config, get_recovery_train_config) inject DR from train/config/gym_config.yaml; eval always uses nominal parameters.

# train/config/gym_config.yaml
domain_randomization:
  m: 0.05         # mass
  I_z: 0.02       # yaw inertia
  lf: 0.05        # CoG fore-distance; lr auto-tracks (wheelbase preserved)
  s_max: 0.03     # max steering angle; s_min auto-mirrors
  sv_max: 0.02    # max steering velocity; sv_min auto-mirrors

For coupled pairs, randomize only the canonical member. The derived partner is recomputed each reset; listing it raises ValueError pointing to the canonical key.

Canonical	Derived	Invariant
`lf`	`lr`	`lf + lr = wheelbase`
`s_max`	`s_min`	`s_min = -s_max`
`sv_max`	`sv_min`	`sv_min = -sv_max`