RoboCAS: A Benchmark for Robotic Manipulation in Complex Object Arrangement Scenarios

Introduction

This repo contains several demonstration trajectories generated by RoboCAS-v0. Details can be found on GitHub.

Dataset Structure

Trajectories are contained in the zip files,each contains the following tasks:

scattered.zip: Picking task in the scattered scene.
orderly.zip: Selecting task in the orderly scene.
stacked.zip: Searching task in the stacked scene.

Details can be found in the paper. Each trajectory is arranged as the following structure:

.
|-- data_info.json                          # Overview of the dataset information, in dictionary format.
|-- episode_0000000                         # Trajectory 0.
|   |-- episode_info.npz                    # Actions and robot states in the trajectory.
|   |-- gripper_camera                      # Data of the camera mounted on the gripper.
|   |   |-- cam_pos_wrt_${parent}.npy       # Camera pose w.r.t. its parent link.
|   |   |-- intrinsic.npy                   # Intrinsic of this camera.
|   |   |-- rgb                             # RGB images in this trajectory.
|   |   |   |-- 0000.png                    # Image at step 0.
|   |   |   |-- ...
|   |   |-- depth                           # Depth images in this trajectory. Same structure as "rgb".
|   |       |-- ...
|   |-- base_camera                         # Data of the camera mounted on the robot base. Same structure as "gripper_camera".
|   |   |-- ...
|   |-- static_camera                       # Data of the camera mounted on the ground. Same structure as "gripper_camera".
|   |   |-- ...
|-- episode_0000001                         # Same structure as "episode_0000000".
|-- ...

In each trajectory folder, the episode_info.npz file contains the trajectory of the agent, the structure and the explanations of each item is as follows. All of the Quaternion in our dataset are stored in "xyzw" order. All relative transforms are calculated under the frame of the last time step, i.e. $\Delta T_{t} = T_{t-1}^{-1} T_{t}$, where $\Delta T_{t}, T_{t} \in SE(3)$ are the relative action and the pose at time step $t$.

rel_pos             # Relative position shift of the EEF w.r.t. the last EEF pose in Cartesian coordinate.
rel_orn             # Relative orientation shift of the EEF w.r.t. the last EEF pose in Quaternion.
ee_pos              # Absolute position of the EEF w.r.t. the arm base in Cartesian coordinate.
ee_orn              # Absolute orientation of the EEF w.r.t. the arm base in Quaternion.
robot_joints        # Joint angles of the arm.
arm_joint_vel       # Velocities of the arm joints.
base_pos            # Position of the arm base w.r.t. the world in Cartesian coordinate.
base_orn            # Orientation of the arm base w.r.t. the world in Quaternion.
base_rel_pos        # Relative position shift of the arm base w.r.t. the the pose in last step in Cartesian coordinate.
base_rel_orn        # Relative orientation shift of the arm base w.r.t. the the pose in last step in Quaternion.
gripper_width       # Open width of the gripper fingers.
gripper_status      # Open/close command of the gripper.
episode_length      # Length of the trajectory.
language_goal       # Global goal instruction of this trajectory.
language_embedding  # Embedding of the goal instruction generated by \href{https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2}{Mini LAMMA}.
step_lang_goals     # Goal annotation of the sub-task for each step of action in this trajectory.
step_goal_embs      # Embedding of step goals generated by \href{https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2}{Mini LAMMA}.
step_goal_type      # Type of the sub-task goals in each step.

Cite

@inproceedings{zheng2024robocas
    title={RoboCAS: A Benchmark for Robotic Manipulation in Complex Object Arrangement Scenarios},
    author={Liming Zheng and Feng Yan and Fanfan Liu and Chengjian Feng and Zhuoliang Kang and Lin Ma},
    year={2024},
    booktitle={NeurIPS: Datasets and Benchmarks Track}
}

image image