LAV2_base_moe

rsl-rl config for LAV2 base task with Mixture-of-Experts trunk.

Mirrors lav2/runner/skrl/cfg/LAV2_base_moe.py but adapted to rsl-rl v5 native dict format. Uses :class:~lav2.runner.rsl_rl.models.moe.MoEMLPModel as the actor model class and :class:~lav2.runner.rsl_rl.algorithms.moe.PPO_MoE for post-update MoE bias correction.

Usage (Isaac Lab Hydra entry point)::

"rsl_rl_cfg_entry_point": "lav2.runner.rsl_rl.cfg.LAV2_base_moe:LAV2MoEPPORunnerCfg"

Usage (direct script)::

from lav2.runner.rsl_rl.cfg.LAV2_base_moe import get_runner_cfg
cfg = get_runner_cfg(experiment_name="my_run")
runner = OnPolicyRunner(env, cfg, log_dir, device=device)

Classes:

Name	Description
BetaDistributionCfg	Configuration for the Beta output distribution.
LAV2MoEPPORunnerCfg	Hydra-compatible config class for the MoE task.

Functions:

Name	Description
get_runner_cfg	Return a rsl-rl v5 native config dict for the MoE task.

BetaDistributionCfg

Configuration for the Beta output distribution.

LAV2MoEPPORunnerCfg

Bases: RslRlOnPolicyRunnerCfg

Hydra-compatible config class for the MoE task.

get_runner_cfg

get_runner_cfg(experiment_name: str = _EXPERIMENT_NAME, max_iterations: int = _MAX_ITERATIONS, num_experts: int = 4, k: int = 2, init_std: float = 1.0, distribution: str = 'beta', action_range: tuple[float, float] = (-1.0, 1.0)) -> dict

Return a rsl-rl v5 native config dict for the MoE task.

Parameters:

Name	Type	Description	Default
experiment_name	str	Logging directory name.	_EXPERIMENT_NAME
max_iterations	int	Number of PPO iterations.	_MAX_ITERATIONS
num_experts	int	Number of experts in the MoE layer.	4
k	int	Top-k experts routed to per token.	2
init_std	float	Initial standard deviation (Gaussian only). Default 1.0 (matches skrl MoE config log_std=0).	1.0
distribution	str	Output distribution, "beta" or "gaussian". Default "beta".	'beta'
action_range	tuple[float, float]	Action space bounds (Beta only). Default (-1.0, 1.0).	(-1.0, 1.0)