Implementation:Isaac sim IsaacGymEnvs AllegroKukaThrow

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

Overview

AllegroKukaThrow is an object throwing task environment where a Kuka arm with an Allegro hand must pick up an object and throw it into a randomly positioned target bucket.

Description

The AllegroKukaThrow class extends AllegroKukaBase to implement a throwing task in the IsaacGym simulation environment. The robotic system must pick up an object from the table, and throw it toward a target bucket placed at a random position in the workspace. The bucket is a physics-enabled asset with collision decomposition (VHACD) that has a fixed base link so it remains stationary during the simulation.

The target bucket is randomized to appear on either the left or right side of the table with additional random offsets in the x, y, and z axes. The goal position is set slightly above the bucket opening (z + 0.05) to account for the trajectory the object must follow. The task uses a single object keypoint at the origin since only the positional accuracy of the throw matters, not the object's orientation during flight.

Like the regrasping task, the object is reset to the table surface when the target is reset, and the lifted_object flag is cleared. The _true_objective method provides a PBT-compatible objective using tolerance_successes_objective to track training progress.

Usage

Use this class when training a reinforcement learning policy for robotic throwing tasks. It is suitable for research into dynamic manipulation, where the agent must learn ballistic release timing and arm/hand coordination to accurately throw objects at variable target locations. Configure it via Hydra configuration files within the IsaacGymEnvs framework.

Code Reference

Source Location

• Repository: IsaacGymEnvs
• File: isaacgymenvs/tasks/allegro_kuka/allegro_kuka_throw.py
• Lines: 41-124

Signature

class AllegroKukaThrow(AllegroKukaBase):
    def __init__(self, cfg, rl_device, sim_device, graphics_device_id, headless, virtual_screen_capture, force_render):
        ...

    def _object_keypoint_offsets(self):
        ...

    def _load_additional_assets(self, object_asset_root, arm_pose):
        ...

    def _create_additional_objects(self, env_ptr, env_idx, object_asset_idx):
        ...

    def _after_envs_created(self):
        ...

    def _reset_target(self, env_ids: Tensor) -> None:
        ...

    def _extra_object_indices(self, env_ids: Tensor) -> List[Tensor]:
        ...

    def _true_objective(self) -> Tensor:
        ...

Import

from isaacgymenvs.tasks.allegro_kuka.allegro_kuka_throw import AllegroKukaThrow

I/O Contract

Inputs

Name	Type	Required	Description
cfg	dict	Yes	Hydra configuration dictionary containing environment parameters, asset paths (including bucket URDF), and tolerance settings
rl_device	str	Yes	Device string for the RL algorithm (e.g., "cuda:0")
sim_device	str	Yes	Device string for the simulation (e.g., "cuda:0")
graphics_device_id	int	Yes	GPU device ID for rendering
headless	bool	Yes	Whether to run without a display
virtual_screen_capture	bool	Yes	Whether to capture virtual screen output
force_render	bool	Yes	Whether to force rendering even in headless mode

Outputs

Name	Type	Description
rew_buf	Tensor	Per-environment reward buffer based on object proximity to the bucket target
is_success	Tensor	Boolean tensor indicating which environments successfully threw the object into the bucket
true_objective	Tensor	PBT objective combining tolerance progress and success count

Usage Examples

# Register and instantiate the AllegroKukaThrow task via IsaacGymEnvs
from isaacgymenvs.tasks.allegro_kuka.allegro_kuka_throw import AllegroKukaThrow

# Typically created through the task registry with Hydra config
env = AllegroKukaThrow(
    cfg=task_cfg,
    rl_device="cuda:0",
    sim_device="cuda:0",
    graphics_device_id=0,
    headless=True,
    virtual_screen_capture=False,
    force_render=False,
)

# Run a step in the environment
obs = env.reset()
actions = policy(obs)
obs, rewards, dones, info = env.step(actions)