SO-ARM101 Reaching Policy
This model is a reinforcement learning policy trained for the SO-ARM101 robot arm to perform end-effector reaching tasks in Isaac Lab.
Model Description
- Task: Move the end-effector to randomly sampled target poses in 3D space
- Robot: SO-ARM101 (6-DOF robotic arm)
- Framework: Isaac Lab 2.3.0 (on Isaac Sim 5.1.0)
- Algorithm: RSL-RL (Robotic Systems Lab - Reinforcement Learning)
- Environment:
Isaac-SO-ARM101-Reach-v0 - Training: 999 iterations with 4096 parallel environments
Model Overview
This policy learns to control the SO-ARM101 robot arm's joint positions to reach target end-effector poses. The model effectively learns inverse kinematics behavior through reinforcement learning, enabling the robot to accurately position its end-effector at desired 3D locations.
Training Details
Environment Configuration
- Observation Space: Joint positions, velocities, and target pose relative to end-effector
- Action Space: Joint position commands (6 DOF)
- Reward Function: Negative distance between end-effector and target pose
- Episode Length: Variable (resets on success or timeout)
Training Parameters
- Parallel Environments: 4096
- Total Iterations: 999
- Training Time: ~1.5 hours on NVIDIA RTX 4080 Super (16GB VRAM)
- Framework: Isaac Lab with RSL-RL runner
- Simulator: Isaac Sim 5.1.0
Hardware Used
- GPU: NVIDIA RTX 4080 Super (16GB VRAM)
- OS: Ubuntu 24.04 LTS
- CUDA: 13.0
Usage
Prerequisites
# Install Isaac Lab (with Docker)
# See: https://isaac-sim.github.io/IsaacLab/
# Clone SO-ARM101 external project
git clone https://github.com/MuammerBay/isaac_so_arm101.git
cd isaac_so_arm101
Evaluation
# Inside Isaac Lab container
cd /workspace/isaaclab
# Run the trained policy
./isaaclab.sh -p /workspace/isaac_so_arm101/src/isaac_so_arm101/scripts/rsl_rl/play.py \
--task Isaac-SO-ARM101-Reach-Play-v0 \
--checkpoint /path/to/model_999.pt
Training From Scratch
# Train the policy
./isaaclab.sh -p /workspace/isaac_so_arm101/src/isaac_so_arm101/scripts/rsl_rl/train.py \
--task Isaac-SO-ARM101-Reach-v0 \
--num_envs 4096 \
--headless
Performance
The trained policy demonstrates accurate reaching behavior with the SO-ARM101 robot, successfully moving the end-effector to target positions across the reachable workspace with high precision.
Use Cases
This reaching policy serves as a foundation for:
- Inverse Kinematics: Learned IK controller for end-effector positioning
- Manipulation Tasks: Base controller for pick-and-place, assembly, etc.
- Trajectory Following: Can be extended for path planning applications
- Sim-to-Real Transfer: Ready for deployment on real SO-ARM101 hardware
Citation
If you use this model, please cite:
@misc{so-arm101-reach-isaaclab,
title={SO-ARM101 Reaching Policy trained with Isaac Lab},
author={PathOn AI},
year={2026},
howpublished={\url{https://huggingface.co/}},
}
@software{isaaclab,
author = {Mittal, Mayank and others},
title = {Isaac Lab: A Unified Framework for Robot Learning},
url = {https://isaac-sim.github.io/IsaacLab/},
year = {2024},
}
Related Resources
License
MIT License