# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md). # All rights reserved. # Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. # # SPDX-License-Identifier: BSD-3-Clause from dataclasses import MISSING from isaaclab.utils.configclass import configclass from isaaclab_rl.rsl_rl import ( RslRlMLPModelCfg, RslRlOnPolicyRunnerCfg, RslRlPpoAlgorithmCfg, ) from isaaclab_tasks.utils import PresetCfg @configclass class BoundedBetaDistributionCfg(RslRlMLPModelCfg.DistributionCfg): """RSL-RL 5.4.1 native bounded distribution with corrected log-probability.""" class_name: str = "BetaDistribution" action_range: tuple[float, float] = (-1.0, 1.0) ## # Model and algorithm presets. ## FIXED_CARTESIAN_POLICY_CFG = RslRlMLPModelCfg( distribution_cfg=BoundedBetaDistributionCfg(), obs_normalization=True, hidden_dims=[256, 256, 128], activation="elu", ) FIXED_CARTESIAN_CRITIC_CFG = RslRlMLPModelCfg( obs_normalization=True, hidden_dims=[256, 256, 128], activation="elu", ) _PPO_HYPERPARAMS = dict( value_loss_coef=1.0, use_clipped_value_loss=True, clip_param=0.2, entropy_coef=0.005, num_learning_epochs=5, num_mini_batches=4, learning_rate=1.0e-4, schedule="adaptive", gamma=0.995, lam=0.90, desired_kl=0.01, max_grad_norm=1.0, ) ALGO_CFG = RslRlPpoAlgorithmCfg(**_PPO_HYPERPARAMS) ## # Runner configurations. ## @configclass class SO101PPOBaseRunnerCfg(RslRlOnPolicyRunnerCfg): num_steps_per_env = 32 max_iterations = 15000 save_interval = 250 experiment_name = (MISSING,) # type: ignore obs_groups = (MISSING,) # type: ignore actor = (MISSING,) # type: ignore critic = (MISSING,) # type: ignore algorithm = MISSING # type: ignore @configclass class SO101FixedCartesianAnchorBenchP0PPORunnerCfg(PresetCfg): """Bounded, physically timed PPO for the full A-Z press/release/clear P0.""" default = SO101PPOBaseRunnerCfg().replace( num_steps_per_env=64, max_iterations=20000, save_interval=500, experiment_name="so101_fixed_cartesian_anchorbench_p0c", obs_groups={"actor": ["policy"], "critic": ["policy"]}, actor=FIXED_CARTESIAN_POLICY_CFG, critic=FIXED_CARTESIAN_CRITIC_CFG, algorithm=ALGO_CFG, # The Beta distribution already has the exact [-1, 1] support. A # wrapper clip would recreate the old off-contract Gaussian behavior. clip_actions=None, ) @configclass class SO101FixedCartesianAnchorBenchP1APPORunnerCfg(PresetCfg): """Direct two-letter P1A hedge using the frozen P0 actor contract.""" default = SO101PPOBaseRunnerCfg().replace( num_steps_per_env=64, max_iterations=20000, save_interval=500, experiment_name="so101_fixed_cartesian_anchorbench_p1a_letters2", obs_groups={"actor": ["policy"], "critic": ["policy"]}, actor=FIXED_CARTESIAN_POLICY_CFG, critic=FIXED_CARTESIAN_CRITIC_CFG, algorithm=ALGO_CFG, clip_actions=None, ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianP0PPORunnerCfg(PresetCfg): """Run-76 PPO contract with an unambiguous calibrated output namespace.""" default = SO101FixedCartesianAnchorBenchP0PPORunnerCfg().default.replace( experiment_name="so101_physical_calibrated_20260718_fixed_cartesian_p0" ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianP1APPORunnerCfg(PresetCfg): """Run-77 PPO contract with an unambiguous calibrated output namespace.""" default = SO101FixedCartesianAnchorBenchP1APPORunnerCfg().default.replace( experiment_name="so101_physical_calibrated_20260718_fixed_cartesian_p1a_letters2" ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianBaselineP1APPORunnerCfg(PresetCfg): """Independent original-Workshop P1A output namespace.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1APPORunnerCfg().default.replace( experiment_name="so101_physical_calibrated_20260718_fixed_cartesian_baseline_p1a_letters2" ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianUSDDriveP1APPORunnerCfg(PresetCfg): """Independent loaded-USD-drive P1A output namespace.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1APPORunnerCfg().default.replace( experiment_name="so101_physical_calibrated_20260718_fixed_cartesian_usd_drive_p1a_letters2" ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianP1BPPORunnerCfg(PresetCfg): """Three-letter continuation of the physically qualified P1A policy.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1APPORunnerCfg().default.replace( experiment_name="so101_physical_calibrated_20260718_fixed_cartesian_p1b_letters3" ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianP1CPPORunnerCfg(PresetCfg): """Four-letter continuation of the physically qualified P1A policy.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1APPORunnerCfg().default.replace( experiment_name="so101_physical_calibrated_20260718_fixed_cartesian_p1c_letters4" ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianP1DPPORunnerCfg(PresetCfg): """Six-letter continuation of the physically qualified P1A policy.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1APPORunnerCfg().default.replace( experiment_name="so101_physical_calibrated_20260718_fixed_cartesian_p1d_letters6" ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianP1DTransit15PPORunnerCfg(PresetCfg): """Six-letter P1D with a 15 mm post-release transit clearance.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1APPORunnerCfg().default.replace( experiment_name=("so101_physical_calibrated_20260718_fixed_cartesian_p1d_transit15_letters6") ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianBaselineP1DTransit15PPORunnerCfg(PresetCfg): """Matched Transit15 continuation using the SO-101 workshop baseline drives.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1DTransit15PPORunnerCfg().default.replace( experiment_name=("so101_physical_calibrated_20260718_fixed_cartesian_baseline_p1d_transit15_letters6") ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianUSDDriveP1DTransit15PPORunnerCfg(PresetCfg): """Loaded-USD-drive Transit15 continuation output namespace.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1DTransit15PPORunnerCfg().default.replace( experiment_name=("so101_physical_calibrated_20260718_fixed_cartesian_usd_drive_p1d_transit15_letters6") ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianGTCBerlinP1APPORunnerCfg(PresetCfg): """GTC Berlin P1A output namespace, distinct from archived profiles.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1APPORunnerCfg().default.replace( experiment_name="so101_physical_calibrated_20260718_fixed_cartesian_gtc_berlin_p1a_letters2" ) @configclass class SO101PhysicalCalibrated20260718FixedCartesianGTCBerlinP1DTransit15PPORunnerCfg(PresetCfg): """GTC Berlin Transit15 continuation output namespace.""" default = SO101PhysicalCalibrated20260718FixedCartesianP1DTransit15PPORunnerCfg().default.replace( experiment_name="so101_physical_calibrated_20260718_fixed_cartesian_gtc_berlin_p1d_transit15_letters6" ) @configclass class SO101GTCBerlinIsaacLabPhrasePPORunnerCfg(PresetCfg): """State-policy PPO for the exact ISAAC LAB 3.0 key sequence.""" default = SO101PhysicalCalibrated20260718FixedCartesianGTCBerlinP1DTransit15PPORunnerCfg().default.replace( max_iterations=5000, save_interval=250, experiment_name="so101_gtc_berlin_state_isaac_lab_3p0", )