Implementation:Isaac sim IsaacGymEnvs AllegroKukaTwoArmsBase

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

Overview

AllegroKukaTwoArmsBase is the foundational VecTask subclass for dual-arm Allegro-Kuka bimanual manipulation tasks, setting up two mirrored KUKA arms with Allegro hands for coordinated two-handed object manipulation.

Description

The AllegroKukaTwoArmsBase class in isaacgymenvs/tasks/allegro_kuka/allegro_kuka_two_arms.py extends VecTask to support bimanual dexterous manipulation. It instantiates two KUKA LBR iiwa arms (7 DOF each) with Allegro hands (16 DOF each), yielding 46 total degrees of freedom. The two arms are positioned symmetrically using configurable X and Y offsets (armXOfs, armYOfs), and the class enforces num_arms == 2.

The observation and action spaces are doubled compared to the single-arm AllegroKukaBase. The observation pipeline in compute_observations() constructs state vectors for both arms including all DOF positions/velocities, fingertip poses for both hands (8 fingertips total), object pose, and goal state. Actions are applied independently to each arm-hand system through pre_physics_step(), supporting either position targets or torque commands for all 46 DOFs.

The reward function compute_kuka_reward() implements keypoint-based rewards that account for both hands' contributions to the manipulation objective. It computes distance deltas from both sets of fingertips to the object, lifting rewards, and coordination bonuses. The class handles synchronized resets of both arm-hand systems and supports domain randomization and curriculum learning inherited from the VecTask infrastructure.

Usage

Use AllegroKukaTwoArmsBase as the base class for bimanual manipulation tasks where two coordinated hands are needed, such as large object manipulation, object handoff between hands, or tasks requiring simultaneous grasping and re-positioning. Subclasses override task-specific reward and reset logic.

Code Reference

Source Location

• Repository: IsaacGymEnvs
• File: isaacgymenvs/tasks/allegro_kuka/allegro_kuka_two_arms.py
• Lines: 1-1415

Signature

class AllegroKukaTwoArmsBase(VecTask):
    def __init__(self, cfg, rl_device, sim_device, graphics_device_id,
                 headless, virtual_screen_capture, force_render):
        """Initialize dual arm+hand+object simulation with 46 DOF control.
        Enforces num_arms == 2 with configurable arm offsets."""

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

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

    def pre_physics_step(self, actions):
        """Apply actions to both arm-hand systems (46 DOFs total)."""

    def post_physics_step(self):
        """Update state tensors for both arms, compute observations and rewards."""

    def compute_observations(self):
        """Build observation: both arms' DOF state, 8 fingertips, object, goals."""

    def compute_kuka_reward(self):
        """Bimanual keypoint reward: distance, lifting, coordination bonuses."""

    def reset_idx(self, env_ids):
        """Reset both arms and object with randomized initial conditions."""

Import

from isaacgymenvs.tasks.allegro_kuka.allegro_kuka_two_arms import AllegroKukaTwoArmsBase

I/O Contract

Inputs

Name	Type	Required	Description
cfg	dict	Yes	Task configuration including env.asset.kukaAllegro (asset path), env.numArms (must be 2), env.armXOfs, env.armYOfs, env.numEnvs, env.episodeLength, reward scale parameters, 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, 46) covering both arms and hands

Outputs

Name	Type	Description
obs_buf	Tensor	Observation tensor containing both arms' DOF positions/velocities, all 8 fingertip poses, object state, and goal information
rew_buf	Tensor	Per-environment scalar rewards combining bimanual distance, lifting, and coordination components
reset_buf	Tensor	Boolean tensor indicating which environments need resetting
extras	dict	Additional info including success flags, consecutive successes, and per-arm reward components

Usage Examples

# AllegroKukaTwoArmsBase is subclassed for specific bimanual tasks.
# Example: Creating a two-arm reorientation task

from isaacgymenvs.tasks.allegro_kuka.allegro_kuka_two_arms import AllegroKukaTwoArmsBase

class AllegroKukaTwoArmsReorientation(AllegroKukaTwoArmsBase):
    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 for bimanual reorientation

    def compute_kuka_reward(self):
        # Override with bimanual coordination reward
        super().compute_kuka_reward()
        # Add handoff or coordinated manipulation bonus

# YAML config must specify:
# env:
#   numArms: 2
#   armXOfs: 0.4
#   armYOfs: 0.3

Related Pages

• Principle:Isaac_sim_IsaacGymEnvs_Dexterous_Manipulation_Task_Design