ganatrask/pika_sim

Pika Sim — Pick & Place Dataset

Simulated pick-and-place dataset generated with MuJoCo using the Agilex Pika gripper model.

Overview

Episodes	500
Steps per episode	149
Success rate	100%
Record frequency	20 Hz
Sim timestep	1 ms
Simulator	MuJoCo ≥ 3.0

Task

A parallel-jaw gripper on a 3-DOF Cartesian gantry picks up a cube from a randomized position and places it at a randomized target location on a table. Positions are randomized per episode (seed 42).

Dataset Structure

dataset/
├── episodes/
│   └── episode_NNNN.h5    # 500 HDF5 files
└── videos/
    └── wrist_cam_NNNN.mp4 # 500 wrist camera videos (640×480, 20fps)

HDF5 Schema

Each episode_NNNN.h5 contains:

├── actions/
│   ├── ee_pose          (149, 7) float32   # target xyz + quaternion (wxyz)
│   └── gripper          (149, 1) float32   # 0.0 = open, 1.0 = closed
├── observations/
│   ├── ee_pose          (149, 7) float32   # current xyz + quaternion (wxyz)
│   ├── gripper_width    (149, 1) float32   # meters, 0.0–0.087
│   └── timestamp        (149,)   float64   # seconds since episode start
├── videos/
│   └── wrist_cam        (149,)   int64     # frame indices into wrist_cam_NNNN.mp4
├── env_state            ()       str/JSON  # pick/place positions, cube final pose, error
├── success              ()       bool
└── episode_length       ()       int64

Conventions

Convention	Value
Action frame	World absolute
Observation frame	World absolute
Quaternion format	wxyz (constant — no rotation DOF)
Gripper action	0 = open, 1 = closed
Gripper max width	0.087 m (87 mm)
Success threshold	< 20 mm placement error

Loading Example

import h5py
import json

with h5py.File("episodes/episode_0000.h5", "r") as f:
    actions_ee = f["actions/ee_pose"][:]        # (149, 7)
    actions_grip = f["actions/gripper"][:]      # (149, 1)
    obs_ee = f["observations/ee_pose"][:]       # (149, 7)
    obs_width = f["observations/gripper_width"][:]  # (149, 1)
    timestamps = f["observations/timestamp"][:] # (149,)
    success = f["success"][()]                  # bool
    env = json.loads(f["env_state"][()])        # dict
    print(f"Pick: {env['pick_pos']}, Place: {env['place_pos']}, Error: {env['placement_error_mm']:.1f}mm")

Generation

cd sim/pika_gripper_mujoco_sim
python record_dataset.py --episodes 500 --seed 42

Source: pika_ros (ros2 branch)