Implementation:Isaac sim IsaacGymEnvs AllegroKukaBase

Knowledge Sources	IsaacGymEnvs
Domains	Robotic_Manipulation, Reinforcement_Learning
Last Updated	2026-02-15 11:00 GMT

Overview

AllegroKukaBase is the foundational VecTask subclass for all single-arm Allegro-Kuka dexterous manipulation tasks, setting up a 7-DOF KUKA arm with a 16-DOF Allegro hand for object manipulation in GPU-parallelized Isaac Gym environments.

Description

The AllegroKukaBase class in isaacgymenvs/tasks/allegro_kuka/allegro_kuka_base.py extends VecTask to provide the complete simulation infrastructure for single-arm dexterous manipulation. It configures a KUKA LBR iiwa robotic arm (7 DOF) combined with an Allegro hand (16 DOF across 4 fingers: index, middle, ring, and thumb), resulting in 23 total degrees of freedom. The class handles asset loading, environment creation with objects (cuboids, sticks, or default cubes), and the full physics simulation loop.

The observation pipeline in compute_observations() constructs a rich state vector including hand DOF positions and velocities, fingertip positions and orientations, object pose and velocity, and relative distances between fingertips and the object. The reward function compute_kuka_reward() implements a keypoint-based reward system that measures the distance between current and goal keypoint configurations, combined with lifting rewards, bonus thresholds, and penalty terms for dropping objects or excessive forces.

The class supports curriculum learning through configurable difficulty scaling, domain randomization via Isaac Gym's built-in randomization system, and privileged actions that allow direct object manipulation for training bootstrapping. Pre-physics and post-physics step methods handle action application (position or torque control), contact force processing, and success/failure detection with automatic environment resets.

Usage

Use AllegroKukaBase as the base class for implementing specific single-arm Allegro-Kuka manipulation tasks such as regrasping, reorientation, or throw-and-catch. Subclasses typically override the reward function, observation space, or reset logic to define task-specific behavior.

Code Reference

Source Location

• Repository: IsaacGymEnvs
• File: isaacgymenvs/tasks/allegro_kuka/allegro_kuka_base.py
• Lines: 1-1592

Signature

class AllegroKukaBase(VecTask):
    def __init__(self, cfg, rl_device, sim_device, graphics_device_id,
                 headless, virtual_screen_capture, force_render):
        """Initialize arm+hand+object simulation with 23 DOF control."""

    def create_sim(self):
        """Create Isaac Gym simulation with ground plane and environments."""

    def _create_envs(self, num_envs, spacing, num_per_row):
        """Load KUKA+Allegro and object assets, create parallel environments."""

    def pre_physics_step(self, actions):
        """Apply actions to arm and hand DOFs (position or torque control)."""

    def post_physics_step(self):
        """Update state tensors, compute observations and rewards, handle resets."""

    def compute_observations(self):
        """Build observation vector: DOF state, fingertips, object pose, goals."""

    def compute_kuka_reward(self):
        """Keypoint-based reward: distance delta + lifting + bonus - penalties."""

    def reset_idx(self, env_ids):
        """Reset specified environments with randomized initial conditions."""

Import

from isaacgymenvs.tasks.allegro_kuka.allegro_kuka_base import AllegroKukaBase

I/O Contract

Inputs

Name	Type	Required	Description
cfg	dict	Yes	Task configuration including env.asset.kukaAllegro (asset path), env.numEnvs, env.episodeLength, reward scale parameters (distanceDeltaRewScale, liftingRewScale, liftingBonus, etc.), and control mode settings
rl_device	str	Yes	Device for RL computations (e.g., "cuda:0")
sim_device	str	Yes	Device for physics simulation (e.g., "cuda:0")
graphics_device_id	int	Yes	GPU ID for rendering (-1 for headless)
headless	bool	Yes	Whether to run without visualization
actions	Tensor	Yes	Per-step action tensor of shape (num_envs, 23) or (num_envs, 26) if privileged actions enabled

Outputs

Name	Type	Description
obs_buf	Tensor	Observation tensor containing DOF positions, velocities, fingertip poses, object state, and goal information
rew_buf	Tensor	Per-environment scalar rewards combining distance, lifting, and bonus components
reset_buf	Tensor	Boolean tensor indicating which environments need resetting
extras	dict	Additional info including success flags, consecutive successes, and reward components

Usage Examples

# AllegroKukaBase is not used directly but subclassed for specific tasks.
# Example: Creating a reorientation task

from isaacgymenvs.tasks.allegro_kuka.allegro_kuka_base import AllegroKukaBase

class AllegroKukaReorientation(AllegroKukaBase):
    def __init__(self, cfg, rl_device, sim_device, graphics_device_id,
                 headless, virtual_screen_capture, force_render):
        super().__init__(cfg, rl_device, sim_device, graphics_device_id,
                         headless, virtual_screen_capture, force_render)
        # Task-specific initialization

    def compute_kuka_reward(self):
        # Override with reorientation-specific reward
        super().compute_kuka_reward()
        # Add orientation alignment bonus

# Register the task:
# isaacgymenvs.task_map["AllegroKukaReorientation"] = AllegroKukaReorientation

Related Pages

• Principle:Isaac_sim_IsaacGymEnvs_Dexterous_Manipulation_Task_Design