nvidia/PhysicalAI-Robotics-mindmap-GR1-Drill-in-Box

Dataset Description:

This dataset is a multimodal collection of trajectories generated in Isaac Lab on the Drill in Box task defined in mindmap.

The task was created to evaluate robot manipulation policies on their spatial memory capabilities. With this (partial) dataset you can generate the full dataset used for mindmap model training, run a mindmap training or evaluate mindmap open/closed loop.

This dataset is for research and development only.

Dataset Owner(s):

NVIDIA Corporation

Dataset Creation Date:

09/26/2025

License/Terms of Use:

This dataset is governed by the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0)

Intended Usage:

This dataset is intended for: Research in robot manipulation policies using imitation learning Training and evaluating robotic manipulation policies on tasks requiring spatial memory Training and evaluating the mindmap model using the mindmap codebase

Dataset Characterization:

Data Collection Method

• Synthetic
• Human teleoperation
• Automatic trajectory generation

20 human teleoperated demonstrations are collected through Apple Vision Pro in Isaac Lab. From these 20 demonstrations, a total of 200 demos are generated automatically using a synthetic motion trajectory generation framework, Isaac Lab Mimic.

Labeling Method

Not Applicable

Dataset Format

We provide the Mimic-generated 200 demonstrations in an HDF5 dataset file, and mindmap-formatted datasets for 10 demonstrations converted from the HDF5 file.

Due to storage limitations, we only provide 10 demonstrations in the mindmap-formatted datasets. If you want to generate the full 200 demos, refer to the mindmap data generation docs.

The mindmap dataset consists of tarred demonstration folders. After untarring, the structure is as follows:

📂 <DATASET_NAME>/
├── 📂 demo_00000/
│   ├── 00000.<CAMERA_NAME>_depth.png
│   ├── 00000.<CAMERA_NAME>_intrinsics.npy
│   ├── 00000.<CAMERA_NAME>_pose.npy
│   ├── 00000.<CAMERA_NAME>_rgb.png
│   ├── 00000.nvblox_vertex_features.zst
│   ├── 00000.robot_state.npy
│   ├── ...
│   ├── <NUM_STEPS_IN_DEMO>.<CAMERA_NAME>_depth.png
│   ├── ...
│   ├── <NUM_STEPS_IN_DEMO>.robot_state.npy
│   └── demo_successful.npy
├── 📂 demo_00001/
│   └── ...
└── 📂 demo_00009>/
└── dataset_name.hdf5

Each demonstration consists of a variable NUM_STEPS with multimodal data:

• RGB-D frame including corresponding camera pose and intrinsics
• Metric-Semantic Reconstruction represented as featurized pointcloud in nvblox_vertex_features.zst
• Robot state including end-effector pose, gripper closedness and head yaw orientation

Dataset Quantification

Record Count: 200 demonstrations/trajectories

Total Storage: 54 GB

Reference(s):

• mindmap paper:
  • Remo Steiner, Alexander Millane, David Tingdahl, Clemens Volk, Vikram Ramasamy, Xinjie Yao, Peter Du, Soha Pouya and Shiwei Sheng. "mindmap: Spatial Memory in Deep Feature Maps for 3D Action Policies". CoRL 2025 Workshop RemembeRL. arXiv preprint arXiv:2509.20297 (2025).
• mindmap codebase:
  • github.com/NVlabs/nvblox_mindmap

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report security vulnerabilities or NVIDIA AI Concerns here.