PolicyIK Data
Paper | Code | Project Page
The data of PolicyIK, a goal-conditioned reinforcement learning policy for high-precision inverse kinematics on the 7-DoF OpenArm. It holds the training pools and evaluation tasks of the simulation experiments, and the evaluation tasks, demonstrations and task pool of the hardware experiments. The released simulation policies are in PolicyIK.
Dataset Structure
The files and paths are identical to those in the code repository. Joint configurations are 7-D, in radians, and positions are in meters.
Simulation
All poses are end-effector poses of openarm_hand in the robot base frame, with quaternions in (w, x, y, z) order.
| Path | Content |
|---|---|
simulation/data/source_set.npz |
200,000 joint configurations q with their poses ee_pos, ee_quat and the floor_contact and self_collision labels |
simulation/data/train_start_pool.npz |
training start configurations (q, ee_pos, ee_quat) |
simulation/data/train_goal_pool.npz |
training goals in the deployment workspace (q, ee_pos, ee_quat) |
simulation/data/eval_tasks.npz |
4,120 evaluation tasks q_start, goal_pos, goal_quat |
Hardware
Each task has a start configuration q_start, a goal position goal_p and a goal rotation matrix goal_R.
| Path | Content |
|---|---|
hardware/data/eval_tasks.json |
60 evaluation tasks, 20 in each start-goal gap bin (20–50 mm, 50–100 mm, over 100 mm) |
hardware/data/demo_tasks.json |
150 large-rotation tasks for demonstration recording |
hardware/data/demos.npz |
22,500 transitions of scripted demonstrations of these tasks recorded on the arm (observation, action, reward, terminated, truncated, next_observation, ep_id, ep_t) |
hardware/data/demos_meta.json |
the action and reward settings of the recording |
hardware/data/task_pool.json |
3,000 long-range tasks that training mixes into its resets |
Download
import json
import numpy as np
from huggingface_hub import snapshot_download
root = snapshot_download("tina390/PolicyIK-Data", repo_type="dataset", revision="v1.0")
sim_tasks = np.load(f"{root}/simulation/data/eval_tasks.npz")
robot_tasks = json.load(open(f"{root}/hardware/data/eval_tasks.json"))["tasks"]
License
Apache-2.0, the same as the code repository. The poses are computed with the OpenArm robot description (openarm_description, Apache-2.0).
Citation
@article{liu2026policyik,
title = {PolicyIK: High-Precision Inverse Kinematics for Robot Arms with Reinforcement Learning},
author = {Liu, Yuhan and Liu, Xinqi and Dong, Zixuan and Ross, Keith W. and Wan, Shenji},
journal = {arXiv preprint arXiv:2611.05372},
year = {2026}
}
- Downloads last month
- 23