ganatrask/food_transfer_ik_dataset

Dataset

Simulated robotic manipulation dataset for food transfer tasks using a Trossen arm in MuJoCo.

Demo

Dataset Description

This dataset contains scripted demonstrations of a 6-DOF Trossen robotic arm performing food transfer tasks - scooping from a source container and dumping into target bowls.

Task Overview

The robot performs a multi-phase food transfer:

1. HOME - Start position
2. APPROACH_CONTAINER - Move above source container
3. REACH_CONTAINER - Lower into container
4. SCOOP - Scooping motion with wrist rotation
5. LIFT - Lift from container
6. APPROACH_BOWL - Move above target bowl
7. LOWER_BOWL - Lower to bowl
8. DUMP - Rotate wrist to dump
9. RETURN - Return home

Dataset Structure

Each episode is stored as an HDF5 file with the following structure:

episode_X.hdf5
├── observations/
│   ├── images/
│   │   ├── main_view    (750, 480, 640, 3) uint8  # Overhead camera
│   │   └── cam          (750, 480, 640, 3) uint8  # Wrist camera
│   ├── qpos             (750, 8) float32          # Joint positions
│   └── qvel             (750, 8) float32          # Joint velocities
├── action               (750, 8) float32          # Joint actions
├── env_state/
│   ├── bowl_1           (7,) float64              # Bowl 1 pose (pos + quat)
│   ├── bowl_2           (7,) float64
│   ├── bowl_3           (7,) float64
│   ├── bowl_4           (7,) float64
│   ├── source_container (7,) float64
│   └── target_container (7,) float64
├── success              () bool
└── attrs:
    ├── sim: True
    ├── source: "food_transfer_ik"
    └── target: <bowl_name>

Robot Configuration

• Robot: Trossen WXAI 6-DOF arm
• Joints 0-5: Arm joints (waist, shoulder, elbow, forearm_roll, wrist_angle, wrist_rotate)
• Joints 6-7: Gripper fingers (coupled)
• End-effector: Spoon attached to wrist

Dataset Statistics

Property	Value
Number of episodes	30
Timesteps per episode	750
Control frequency	50 Hz
Image resolution	480 x 640
Camera views	2 (overhead + wrist)

Usage

import h5py
import numpy as np

with h5py.File("episode_0.hdf5", "r") as f:
    images_overhead = f["observations/images/main_view"][:]
    images_wrist = f["observations/images/cam"][:]
    qpos = f["observations/qpos"][:]
    actions = f["action"][:]
    target_bowl = f.attrs["target"]