datasets:
- nvidia/PhysicalAI-Robotics-mindmap-Franka-Cube-Stacking
- nvidia/PhysicalAI-Robotics-mindmap-Franka-Mug-in-Drawer
- nvidia/PhysicalAI-Robotics-mindmap-GR1-Drill-in-Box
- nvidia/PhysicalAI-Robotics-mindmap-GR1-Stick-in-Bin
Model Overview
Description:
mindmap is a 3D diffusion policy that generates robot trajectories based on a semantic 3D reconstruction of the environment,
enabling robots with spatial memory.
Trained models are available on Hugging Face: PhysicalAI-Robotics-mindmap-Checkpoints
License/Terms of Use
Deployment Geography:
Global
Use Case
The trained mindmap policies allow for quick evaluation of the mindmap concept on selected simulated robotic manipulation tasks.
- Researchers, Academics, Open-Source Community: AI-driven robotics research and algorithm development.
- Developers: Integrate and customize AI for various robotic applications.
- Startups & Companies: Accelerate robotics development and reduce training costs.
References(s):
mindmappaper:- 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).
mindmapcodebase:
Model Architecture:
Architecture Type: Denoising Diffusion Probabilistic Model
Network Architecture:
mindmap is a Denoising Diffusion Probabilistic Model that samples robot trajectories conditioned on sensor observations and a 3D reconstruction of the environment. Images are first passed through a Vision Foundation Model and then back-projected, using the depth image, to a pointcloud. In parallel, a reconstruction of the scene is built that accumulates metric-semantic information from past observations. The two 3D data sources, the instantaneous visual observation and the reconstruction, are passed to a transformer that iteratively denoises robot trajectories.
This model was developed based on: 3D Diffuser Actor
Number of model parameters: ∼3M trainable, plus ∼100M frozen in the image encoder
Input:
Input Type(s):
- RGB: Image frames
- Geometry: Depth frames converted to 3D pointclouds
- State: Robot proprioception
- Reconstruction: Metric-semantic reconstruction represented as featurized pointcloud
Input Format(s):
- RGB: float32 in the range
[0, 1] - Geometry: float32 in world coordinates
- State: float32 in world coordinates
- Reconstruction (represented as feature pointcloud):
- Points: float32 in world coordinates
- Features: float32
Input Parameters:
- RGB:
[NUM_CAMERAS, 3, HEIGHT, WIDTH]-512x512resolution on the provided checkpoints - Geometry:
[NUM_CAMERAS, 3, HEIGHT, WIDTH]-512x512resolution on the provided checkpoints - State:
[HISTORY_LENGTH, NUM_GRIPPERS, 8]- consisting of end-effector translation, rotation (quaternion, wxyz) and closedness - Reconstruction (represented as feature pointcloud):
- Points:
[NUM_POINTS, 3]-NUM_POINTSis 2048 for the provided checkpoints - Features:
[NUM_POINTS, FEATURE_DIM]-FEATURE_DIMis 768 for theRADIO_V25_Bfeature extractor used for the provided checkpoints
- Points:
Output:
Output Type(s): Robot actions
Output Format: float32
Output Parameters:
- Gripper:
[PREDICTION_HORIZON, NUM_GRIPPERS, 8]- consisting of end-effector translation, rotation (quaternion, wxyz) and closedness - Head Yaw:
[PREDICTION_HORIZON, 1]- only for humanoid embodiments
Software Integration:
Runtime Engine(s): PyTorch
Supported Hardware Microarchitecture Compatibility:
- NVIDIA Ampere
- NVIDIA Blackwell
- NVIDIA Jetson
- NVIDIA Hopper
- NVIDIA Lovelace
- NVIDIA Pascal
- NVIDIA Turing
- NVIDIA Volta
Preferred/Supported Operating System(s):
- Linux
Model Version(s):
This is the initial version of the model, version 1.0.0
Training, Testing, and Evaluation Datasets:
Datasets:
- cube_stacking_checkpoint: Franka Cube Stacking Dataset
- mug_in_drawer_checkpoint: Franka Mug in Drawer Dataset
- drill_in_box_checkpoint: GR1 Drill in Box Dataset
- stick_in_bin_checkpoint: GR1 Stick in Bin Dataset
The models were trained on 100 (GR1) and 130 (Franka) demonstrations. The evaluation set consisted of 20 distinct demonstrations. Closed loop testing was performed on 100 demonstrations mutually exclusive from the training set.
Inference:
Engine: PyTorch
Test Hardware: Linux, L40S
Model Limitations:
This model is not tested or intended for use in mission critical applications that require functional safety. The use of the model in those applications is at the user's own risk and sole responsibility, including taking the necessary steps to add needed guardrails or safety mechanisms.
- Risk: This policy is only effective on the exact simulation environment it was trained on.
- Mitigation: Need to retrain the model on new simulation environments.
- Risk: The policy was never tested on a physical robot and likely only works in simulation
- Mitigation: Expand training, testing and validation on physical robot platforms.
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.
For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards.
Please report security vulnerabilities or NVIDIA AI Concerns here.
Bias
| Field | Response |
|---|---|
| Participation considerations from adversely impacted groups protected classes in model design and testing: | Not Applicable |
| Bias Metric (If Measured): | Not Applicable |
| (For GPAI Models) Which characteristic (feature) show(s) the greatest difference in performance?: | Not Applicable |
| (For GPAI Models): Which feature(s) have have the worst performance overall? | Not Applicable |
| Measures taken to mitigate against unwanted bias: | Not Applicable |
| (For GPAI Models): If using internal data, description of methods implemented in data acquisition or processing, if any, to address the prevalence of identifiable biases in the training, testing, and validation data: | Not Applicable |
| (For GPAI Models): Tools used to assess statistical imbalances and highlight patterns that may introduce bias into AI models: | Not Applicable |
| (For GPAI Models): Tools used to assess statistical imbalances and highlight patterns that may introduce bias into AI models: | Not Applicable |
Explainability
| Field | Response |
|---|---|
| Intended Task/Domain: | Robotic Manipulation |
| Model Type: | Denoising Diffusion Probabilistic Model |
| Intended Users: | Roboticists and researchers in academia and industry who are interested in robot manipulation research |
| Output: | Actions consisting of end-effector poses, gripper states and head orientation. |
| (For GPAI Models): Tools used to evaluate datasets to identify synthetic data and ensure data authenticity. | Not Applicable |
| Describe how the model works: | mindmap is a Denoising Diffusion Probabilistic Model that samples robot trajectories conditioned on sensor observations and a 3D reconstruction of the environment. |
| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Not Applicable |
| Technical Limitations & Mitigation: | - The policy is only effective on the exact simulation environment it was trained on. - The policy was never tested on a physical robot and likely only works in simulation. |
| Verified to have met prescribed NVIDIA quality standards: | Yes |
| Performance Metrics: | Closed loop success rate on simulated robotic manipulation tasks. |
| Potential Known Risks: | The model might be susceptible to rendering changes on the simulation tasks it was trained on. |
| Licensing: | NVIDIA Open Model License Agreement |
Safety and Security
| Field | Response |
|---|---|
| Model Application Field(s): | Robotics |
| Describe the life critical impact (if present). | Not Applicable |
| (For GPAI Models): Description of methods implemented in data acquisition or processing, if any, to address other types of potentially harmful data in the training, testing, and validation data: | Not GPAI |
| (For GPAI Models): Description of any methods implemented in data acquisition or processing, if any, to address illegal or harmful content in the training data, including, but not limited to, child sexual abuse material (CSAM) and non-consensual intimate imagery (NCII) | Not GPAI |
| Use Case Restrictions: | Abide by NVIDIA Open Model License Agreement |
| Model and dataset restrictions: | The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development. Restrictions enforce dataset access during training, and dataset license constraints adhered to. |
Privacy
| Field | Response |
|---|---|
| Generatable or reverse engineerable personal data? | No |
| Personal data used to create this model? | No |
| Was consent obtained for any personal data used? | Not Applicable |
| (For GPAI Models): A description of any methods implemented in data acquisition or processing, if any, to address the prevalence of personal data in the training data, where relevant and applicable. | Not Applicable |
| How often is dataset reviewed? | Before Release |
| Is there provenance for all datasets used in training? | Yes |
| Does data labeling (annotation, metadata) comply with privacy laws? | Yes |
| Is data compliant with data subject requests for data correction or removal, if such a request was made? | Yes |
| Applicable Privacy Policy | NVIDIA Privacy Policy |