Implementation:Isaac sim IsaacGymEnvs AllegroKukaUtils

Knowledge Sources	IsaacGymEnvs
Domains	Robotic_Control, Curriculum_Learning
Last Updated	2026-02-15 11:00 GMT

Overview

AllegroKukaUtils provides a DofParameters dataclass and utility functions for configuring joint properties and managing curriculum-based tolerance schedules in Allegro-Kuka manipulation tasks.

Description

This module contains the DofParameters dataclass and several utility functions that are shared across all Allegro-Kuka task variants (regrasping, reorientation, throwing, and their two-arm counterparts). The DofParameters dataclass encapsulates the stiffness, damping, effort, friction, and armature parameters for both the Kuka arm DOFs and the Allegro hand DOFs, with a factory method from_cfg that extracts these values from a Hydra configuration dictionary.

The populate_dof_properties function applies DofParameters values to the IsaacGym DOF property arrays, setting separate stiffness, effort, damping, friction, and armature values for the arm and hand joints. This ensures consistent joint configuration across all environments and task variants.

The tolerance_curriculum function implements a curriculum learning strategy that progressively decreases the success tolerance threshold. It checks whether enough training frames have passed since the last update and whether the policy's mean successes per episode exceed a threshold (3.0) before tightening the tolerance by a configurable increment. The tolerance_successes_objective function computes a PBT-compatible objective that interpolates between 0 (initial tolerance) and 1 (target tolerance) and adds weighted success counts, ensuring that tolerance improvements always outweigh raw success gains at earlier curriculum stages.

Usage

Use DofParameters and populate_dof_properties when setting up joint configurations for Allegro-Kuka environments. Use tolerance_curriculum and tolerance_successes_objective to implement curriculum learning and PBT objectives in custom task subclasses. These utilities are called by all task variants in the allegro_kuka package.

Code Reference

Source Location

• Repository: IsaacGymEnvs
• File: isaacgymenvs/tasks/allegro_kuka/allegro_kuka_utils.py
• Lines: 38-158

Signature

@dataclass
class DofParameters:
    allegro_stiffness: float
    kuka_stiffness: float
    allegro_effort: float
    kuka_effort: List[float]
    allegro_damping: float
    kuka_damping: float
    dof_friction: float
    allegro_armature: float
    kuka_armature: float

    @staticmethod
    def from_cfg(cfg: Dict) -> DofParameters:
        ...

def populate_dof_properties(hand_arm_dof_props, params: DofParameters, arm_dofs: int, hand_dofs: int) -> None:
    ...

def tolerance_curriculum(
    last_curriculum_update: int,
    frames_since_restart: int,
    curriculum_interval: int,
    prev_episode_successes: Tensor,
    success_tolerance: float,
    initial_tolerance: float,
    target_tolerance: float,
    tolerance_curriculum_increment: float,
) -> Tuple[float, int]:
    ...

def tolerance_successes_objective(
    success_tolerance: float, initial_tolerance: float, target_tolerance: float, successes: Tensor
) -> Tensor:
    ...

Import

from isaacgymenvs.tasks.allegro_kuka.allegro_kuka_utils import DofParameters, tolerance_curriculum

I/O Contract

Inputs

Name	Type	Required	Description
cfg	Dict	Yes	Hydra configuration dictionary (for DofParameters.from_cfg) containing keys like allegroStiffness, kukaStiffness, etc.
hand_arm_dof_props	numpy.ndarray	Yes	Structured array of DOF properties from IsaacGym (for populate_dof_properties)
params	DofParameters	Yes	Joint parameter dataclass instance (for populate_dof_properties)
arm_dofs	int	Yes	Number of arm degrees of freedom (for populate_dof_properties)
hand_dofs	int	Yes	Number of hand degrees of freedom (for populate_dof_properties)
last_curriculum_update	int	Yes	Frame number of the last curriculum update (for tolerance_curriculum)
frames_since_restart	int	Yes	Total frames since training restart (for tolerance_curriculum)
curriculum_interval	int	Yes	Minimum frames between curriculum updates (for tolerance_curriculum)
prev_episode_successes	Tensor	Yes	Per-environment success counts from the previous episode (for tolerance_curriculum)
success_tolerance	float	Yes	Current success tolerance threshold (for tolerance_curriculum and tolerance_successes_objective)
initial_tolerance	float	Yes	Starting tolerance value (for tolerance_curriculum and tolerance_successes_objective)
target_tolerance	float	Yes	Final target tolerance value (for tolerance_curriculum and tolerance_successes_objective)
tolerance_curriculum_increment	float	Yes	Multiplicative factor applied to tolerance on each curriculum step (for tolerance_curriculum)
successes	Tensor	Yes	Per-environment success count tensor (for tolerance_successes_objective)

Outputs

Name	Type	Description
DofParameters	DofParameters	Dataclass instance containing arm and hand joint parameters
(success_tolerance, last_curriculum_update)	Tuple[float, int]	Updated tolerance and frame counter from tolerance_curriculum
true_objective	Tensor	PBT objective tensor from tolerance_successes_objective

Usage Examples

from isaacgymenvs.tasks.allegro_kuka.allegro_kuka_utils import (
    DofParameters, populate_dof_properties, tolerance_curriculum, tolerance_successes_objective
)

# Create DOF parameters from config
dof_params = DofParameters.from_cfg(cfg)

# Apply to IsaacGym DOF properties
populate_dof_properties(hand_arm_dof_props, dof_params, arm_dofs=7, hand_dofs=16)

# Run curriculum update
new_tolerance, new_last_update = tolerance_curriculum(
    last_curriculum_update=0,
    frames_since_restart=10000,
    curriculum_interval=5000,
    prev_episode_successes=successes_tensor,
    success_tolerance=0.1,
    initial_tolerance=0.5,
    target_tolerance=0.01,
    tolerance_curriculum_increment=0.9,
)

# Compute PBT objective
objective = tolerance_successes_objective(
    success_tolerance=new_tolerance,
    initial_tolerance=0.5,
    target_tolerance=0.01,
    successes=successes_tensor,
)