[Question] Clarification on Domain Randomization Distributions in DWL

[hesic73]

In the Humanoid Gym environment, the type of distribution (uniform or Gaussian) used for each domain randomization term is clearly specified, as shown below:

However, in the DWL paper, the distribution type for each randomized parameter isn’t mentioned. Could you please clarify whether each term is sampled from a uniform or Gaussian distribution?

Thanks in advance!

[QianZhao-eng]

Both sampling strategies offer distinct trade-offs. Uniform distributions ensure equal exploration probability across the entire action space, which can improve robustness in high-dimensional whole-body control. Gaussian distributions concentrate sampling around the policy mean while allowing tail exploration, producing smoother, more natural motion trajectories essential for continuous motor control.
Consequently, DWL uses uniform sampling for its balanced exploration properties, while humanoid-gym employs Gaussian sampling to generate smoother action sequences and leverage natural stochasticity for exploration noise.