feat: add OneRobotics A1 assets and Reach environments

[T1Amoo]

Description

Adds the OneRobotics A1 robotic arm assets/configurations and contributed Reach environments discussed in #7095.

This contribution includes two supported A1 configurations:

• the fixed-base 7-DoF OneRobotics A1 right arm;
• the fixed-base 14-DoF OneRobotics A1 bimanual robot with stand.

It also adds two Reach environments:

• IsaacContrib-Reach-OneRobotics-A1
• IsaacContrib-Reach-OneRobotics-A1-Bimanual

Both tasks reuse Isaac Lab's current Reach framework and provide RSL-RL PPO configurations.

The unimanual A1 remains a reusable standalone robot asset. In the Reach environment it is mounted directly on the standard Isaac Lab Reach table, similar to a conventional table-mounted manipulator, without introducing an additional pedestal or combined robot/table model.

The bimanual configuration uses the authoritative hardware-provided whole-robot URDF, including its stand and fixed shoulder mounts.

A small A1-specific MDP component generates kinematically reachable pose commands from the source joint limits, joint axes, and fixed transforms. The FK implementation is regression-tested against the spawned end-effector poses in simulation.

Related proposal: #7095

Asset source and licensing

The review-stage robot assets are retrieved from the public OneRobotics asset repository:

https://github.com/katazen/onerobot_h1

An immutable review bundle is also available as:

onerobotics-a1-2026-review-v1

The OneRobotics A1 URDF and mesh assets are licensed under CC BY 4.0.

The Isaac Lab integration code follows the repository's BSD-3-Clause contribution conventions.

NVIDIA legal approval and the final hosted asset URI are still pending. The current Git-backed asset loading path is intended for technical review and can be replaced with the final approved hosted asset path once that process is complete.

Implementation details

Unimanual A1

• 7 actuated joints
• fixed-base configuration
• standard Isaac Lab Reach table
• 7-D joint-position action space
• Link7 end effector
• FK-reachable pose targets
• 200 Hz simulation / 50 Hz policy control
• RSL-RL PPO configuration

The robot base placement is derived from the table geometry and A1 base geometry rather than visually approximated.

A deterministic sample of 4096 FK targets showed 98.56% of target end-effector positions above the tabletop, so the existing reachable-target distribution was preserved without adding table-specific rejection sampling.

Bimanual A1

• 14 actuated joints, ordered right arm then left arm
• fixed-base whole-robot model with stand
• independent right/left 7-D joint-position actions
• Link_r7 and Link_l7 end effectors
• independent FK-reachable targets for both arms
• simultaneous right-and-left success criterion
• 56-D policy observation
• RSL-RL PPO configuration

The bimanual FK preserves the source fixed shoulder transforms and arbitrary joint axes and is checked against the corresponding simulated end-effector poses.

Validation

The final contribution was validated with:

• ./isaaclab.sh --format
• git diff --check
• 7/7 focused A1 asset tests
• 2/2 unimanual Reach tests
• 3/3 bimanual Reach tests
• environment discovery for both task IDs
• public asset retrieval and source-fidelity checks
• zero-action and random-action runtime checks
• exact FK-to-sim end-effector regression checks

GPU finite-state validation:

• unimanual: 256 environments × 500 steps
• bimanual: 256 environments × 500 steps
• 128,000 transitions per task
• observations, rewards, joint states, commands, and applied efforts remained finite

Short RSL-RL integration validation:

• unimanual: 256 environments × 5 iterations
• bimanual: 256 environments × 5 iterations
• 30,720 transitions per task
• finite rollout / optimization metrics
• checkpoints generated successfully

These short runs validate the training integration path only; they are not intended as convergence or policy-quality claims.

No mass, inertia, collision geometry, visual geometry, kinematic transform, joint axis, or joint-limit changes were introduced to make training easier. Actuator limits and drive parameters are overlaid from confirmed OneRobotics hardware data because the CAD-exported URDF contains placeholder effort/velocity fields.

Type of change

• New feature (non-breaking change which adds functionality)
• Documentation update

Release backport

• Backport this pull request to the active release branch after it merges into develop

Screenshots

OneRobotics A1 unimanual Reach

OneRobotics A1 bimanual Reach

Checklist

Docker and GPU tests run on demand. Push the commits you want tested, then
comment run-ci on the pull request.

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my feature works
• I have added a changelog fragment under source/<pkg>/changelog.d/ for every touched package
• I have added my name to the CONTRIBUTORS.md or my name already exists there