Ekshan267/franka-place-cube-in-box-mimic-dataset

Franka Place Cube Into Box — Isaac Lab Mimic Dataset

Dataset Summary

This dataset contains robot manipulation demonstrations for a pick-and-place task where a Franka Panda robot picks up a cube and places it inside an open box. Demonstrations were collected in Isaac Lab via keyboard teleoperation and augmented to 258 demonstrations using Isaac Lab Mimic with CuRobo motion planning.

Each demonstration includes RGB observations from two cameras (wrist-mounted + fixed overhead), proprioceptive state, and action data — making the dataset suitable for training visuomotor policies such as ACT or Diffusion Policy.

---

Files

File	Size	Demos	Description
source_demos.hdf5	307 MB	14	Raw keyboard-teleoperated demonstrations
annotated_demos.hdf5	280 MB	14	Source demos annotated with subtask signals and datagen info
pick_n_place_258.hdf5	5.13 GB	258	Mimic-generated demonstrations with full camera observations

---

Task Description

• Robot: Franka Panda (7-DOF arm + parallel gripper)
• Task: Pick up a cube from the table and place it inside an open box
• Subtasks:
  1. Grasp — reach the cube, close the gripper, and lift off the table (> 15 cm height)
  2. Place — carry the cube to the box and release it inside
• Success condition: Cube is geometrically inside the box (within 9 cm half-width in XY, correct Z range)
• Simulation: Isaac Lab, dt = 1/60s, decimation = 2, DLAA antialiasing

---

Dataset Statistics

	source_demos.hdf5	annotated_demos.hdf5	pick_n_place_258.hdf5
Num demos	14	14	258
Min episode length	179	179	225
Max episode length	502	502	333
Mean episode length	272.4	272.4	232.9
Total steps	3,814	3,814	60,078

---

Observation Space

source_demos.hdf5 and pick_n_place_258.hdf5

Key	Shape	dtype	Description
joint_pos	(T, 9)	float32	Robot joint positions (7 arm + 2 finger)
joint_vel	(T, 9)	float32	Robot joint velocities
eef_pos	(T, 3)	float32	End-effector position in robot base frame
eef_quat	(T, 4)	float32	End-effector orientation (quaternion) in robot base frame
cube_pos	(T, 3)	float32	Cube position in robot base frame
box_pos	(T, 3)	float32	Box position in robot base frame
table_cam	(T, 256, 256, 3)	uint8	RGB from fixed overhead table camera
wrist_cam	(T, 128, 128, 3)	uint8	RGB from wrist-mounted camera

Additional keys in annotated_demos.hdf5

Key	Shape	dtype	Description
datagen_info/eef_pose/panda_hand	(T, 4, 4)	float32	EEF pose as 4×4 homogeneous transform
datagen_info/object_pose/cube	(T, 4, 4)	float32	Cube pose as 4×4 homogeneous transform
datagen_info/object_pose/box	(T, 4, 4)	float32	Box pose as 4×4 homogeneous transform
datagen_info/target_eef_pose/panda_hand	(T, 4, 4)	float32	Target EEF pose as 4×4 homogeneous transform
datagen_info/subtask_start_signals/grasp	(T, 1)	float32	Binary signal: start of grasp subtask
datagen_info/subtask_start_signals/place	(T,)	float32	Binary signal: start of place subtask
datagen_info/subtask_term_signals/grasp	(T,)	float32	Binary signal: termination of grasp subtask
datagen_info/subtask_term_signals/place	(T,)	float32	Binary signal: termination of place subtask

---

Action Space

Key	Shape	dtype	Description
actions	(T, 7)	float32	Differential IK pose commands (6 DOF relative) + binary gripper
processed_actions	(T, 8)	float32	Processed actions stored by Isaac Lab recorder

---

Camera Setup

Table Camera — fixed overhead

• Resolution: 256 × 256
• World position: (2.0, 0.0, 1.2)
• Orientation: rot=(0.354, -0.612, -0.612, 0.354) ROS convention
• Focal length: 24 mm, horizontal aperture: 20.955 mm
• Clipping range: 0.1–10.0 m

Wrist Camera — mounted on panda_hand

• Resolution: 128 × 128
• Offset from hand: pos=(0.13, 0.0, -0.15), rot=(-0.706, 0.037, 0.037, -0.706) ROS convention
• Focal length: 24 mm, horizontal aperture: 20.955 mm
• Clipping range: 0.1–2.0 m

---

Scene Configuration

Element	Details
Table	Seattle Lab Table (Isaac Nucleus)
Cube	DexCube (scale 0.8), randomized: x∈[0.4, 0.5], y∈[-0.1, 0.1], z=0.055
Box	Isaac Lab Nucleus box (scale 0.8), randomized: x∈[0.6, 0.7], y∈[0.15, 0.25], z=0.055
Controller	Differential IK (relative pose, DLS), scale=0.5
Gripper	Binary joint position (open=0.04, close=0.0)

---

Data Generation Pipeline

Step 1 — Teleoperation → source_demos.hdf5

14 demonstrations collected via keyboard teleoperation using Isaac Lab's record_demos.py. Cube and box positions are randomized each episode.

Step 2 — Annotation → annotated_demos.hdf5

Source demos annotated with subtask signals using annotate_demos.py:

• grasp signal: 1 when gripper width < 0.05, EEF within 8 cm of cube, and cube lifted > 15 cm
• place signal: 1 when cube is geometrically inside the box

Step 3 — Mimic Generation → pick_n_place_258.hdf5

Augmented to 258 demos using Isaac Lab Mimic + CuRobo motion planning:

Subtask 0 — grasp:
  object_ref:               cube
  subtask_term_offset_range: (10, 20)   # extend skill 10-20 frames past grasp signal
  subtask_start_offset_range: (0, 0)
  selection_strategy:       nearest_neighbor_object (k=3)
  num_interpolation_steps:  5

Subtask 1 — place:
  object_ref:               box
  subtask_term_offset_range: (0, 0)
  subtask_start_offset_range: (22, 22)  # shift start forward to satisfy strict < constraint
  selection_strategy:       nearest_neighbor_object (k=3)
  num_interpolation_steps:  5

Note on offsets: In the source data, grasp_term and place_start signals fire at the same timestep (gap = 0). Isaac Lab Mimic internally adds +1 to end_index for array indexing, so the non-overlap constraint becomes start_offset_min > term_offset_max + 1. With term_offset_max = 20, the minimum valid start_offset_min is 22.

---

HDF5 Structure

data/
  demo_0/
    actions                          (T, 7)   float32
    processed_actions                (T, 8)   float32
    obs/
      joint_pos                      (T, 9)   float32
      joint_vel                      (T, 9)   float32
      eef_pos                        (T, 3)   float32
      eef_quat                       (T, 4)   float32
      cube_pos                       (T, 3)   float32
      box_pos                        (T, 3)   float32
      table_cam                      (T, 256, 256, 3)  uint8
      wrist_cam                      (T, 128, 128, 3)  uint8
    initial_state/
      articulation/robot/            joint_position, joint_velocity, root_pose, root_velocity
      rigid_object/cube/             root_pose, root_velocity
      rigid_object/box/              root_pose, root_velocity
    states/
      articulation/robot/            joint_position, joint_velocity, root_pose, root_velocity
      rigid_object/cube/             root_pose, root_velocity
      rigid_object/box/              root_pose, root_velocity
  demo_1/
    ...

---

Usage

import h5py
import numpy as np

with h5py.File("pick_n_place_258.hdf5", "r") as f:
    demos = list(f["data"].keys())
    print(f"Total demos: {len(demos)}")  # 258

    demo = demos[0]
    table_cam  = f[f"data/{demo}/obs/table_cam"][:]   # (T, 256, 256, 3) uint8
    wrist_cam  = f[f"data/{demo}/obs/wrist_cam"][:]   # (T, 128, 128, 3) uint8
    joint_pos  = f[f"data/{demo}/obs/joint_pos"][:]   # (T, 9)  float32
    actions    = f[f"data/{demo}/actions"][:]          # (T, 7)  float32

---

License

Apache 2.0