oscarqjh/EB-Manipulation_easi

EB-Manipulation Dataset

EB-Manipulation is a benchmark for evaluating LLM-controlled robotic manipulation in CoppeliaSim using a Franka Panda arm with a parallel gripper. It is part of the EmbodiedBench benchmark suite, designed for use with the EASI evaluation framework.

Dataset Description

Agents must output sequences of 7D discrete gripper actions [X, Y, Z, Roll, Pitch, Yaw, Gripper] to complete manipulation tasks (picking, stacking, placing, wiping). The benchmark tests spatial reasoning, visual understanding, common sense, and complex instruction following.

Subsets

Subset	Description	Episodes
base	Standard manipulation tasks with direct instructions	48
common_sense	Tasks requiring commonsense reasoning about objects	48
complex	Complex multi-step manipulation instructions	48
spatial	Tasks with relative spatial references	48
visual	Tasks requiring visual property recognition	36

Task Types

Task Type	Base Task	Description
pick	pick_cube	Pick up a target object and place it into a container
stack	stack_cubes	Stack cubes in a specified order
place	place_into_shape_sorter	Place objects into the correct shape sorter slots
wipe	wipe_table	Wipe a specified area on the table

Action Space

Each action is a 7D discrete array: [X, Y, Z, Roll, Pitch, Yaw, Gripper_state]

• X, Y, Z: 3D position in voxel grid (range [0, 100])
• Roll, Pitch, Yaw: Discrete Euler angles (range [0, 120], each unit = 3 degrees)
• Gripper state: 0 = close, 1 = open

Dataset Structure

.
├── data/
│   ├── base.jsonl
│   ├── common_sense.jsonl
│   ├── complex.jsonl
│   ├── spatial.jsonl
│   └── visual.jsonl
├── simulator_data.zip          # Binary simulation files (auto-extracted by EASI)
│   ├── data/                   # Per-split episode data (.ttm, .pkl)
│   ├── vlm/                    # Task templates and object models
│   └── amsolver/robot_ttms/    # Robot model files
└── README.md

Data Fields (JSONL)

Each row in the JSONL files contains:

• id (int): Unique identifier within the split
• task_name (string): Task variation name (e.g., pick_cube_shape)
• variation (int): Variation number within the task
• episode_num (int): Episode number within the variation
• instruction (string): Natural language task instruction
• task_type (string): Base task type (pick, stack, place, wipe)

Simulator Data (simulator_data.zip)

Each episode's binary data is stored at: data/{split}/eval/{task_name}/variation{N}/episodes/episode{N}/

• task_base.ttm — CoppeliaSim scene state
• waypoint_sets.ttm — Waypoint configuration
• configs.pkl — Episode metadata and success conditions

Usage

Loading with Datasets Library

from datasets import load_dataset

# Load a specific split
dataset = load_dataset("oscarqjh/EB-Manipulation_easi", split="base")

# Access data
for example in dataset:
    print(example["instruction"])
    print(example["task_name"])

Using with EASI

# Run evaluation on the base split
easi run ebmanipulation_base --agent react --backend openai --model gpt-4o

# List available manipulation splits
easi task list | grep ebmanipulation

Requirements

• CoppeliaSim V4.1.0
• PyRep (CoppeliaSim Python binding)
• AMSolver (modified RLBench fork)

Acknowledgements

This dataset is derived from the EmbodiedBench EB-Manipulation benchmark and uses CoppeliaSim as the simulation environment.