PhysicalAI Robotics Kitchen Manipulation Demos
PhysicalAI Robotics Kitchen Manipulation Demos is a large-scale dataset of 600 hours of human-teleoperated demonstrations across 316 different kitchen manipulation tasks, totalling 55,000 trajectories collected using a Franka Panda robot mounted on an Omron mobile base. Released in March 2026 under CC BY 4.0 for commercial use, the dataset is the largest kitchen-specific robot manipulation dataset collected on real hardware. The 316-task coverage spans the full range of kitchen manipulation operations: food preparation, cooking, cleaning, appliance operation, and object organisation. The mobile base configuration captures whole-body kitchen navigation alongside arm manipulation — enabling training of robots that must move through the kitchen to complete multi-step recipes. At 55K trajectories it is 10x larger than the next largest kitchen-specific real-robot dataset.
| Year | 2026 |
|---|---|
| Total hours | 600 |
| Embodiments | Franka Panda, Omron mobile base |
| Modalities | rgb, proprioception, depth |
| Task categories | manipulation, cooking, cleaning, pick-and-place, long-horizon |
| Data format | hdf5, parquet, mp4 |
| License | CC BY 4.0 |
| Access | open — commercial use permitted |
| Maintainer | NVIDIA |
| Origin country | US |
What is it?
PhysicalAI Robotics Kitchen Manipulation Demos is a dataset of 600 hours of human-teleoperated demonstrations across 316 different kitchen manipulation tasks, totalling 55,000 trajectories collected using a Franka Panda robot on an Omron mobile base. Released in March 2026 under CC BY 4.0, the mobile base configuration captures whole-body kitchen navigation alongside arm manipulation, enabling training of robots that must move through the kitchen to complete multi-step recipes. At 55K trajectories it is 10x larger than the next largest kitchen-specific real-robot dataset.
Who is it for?
Researchers and companies working on kitchen robot deployment — one of the highest-priority markets for service robots. The 316-task coverage spans the full spectrum of kitchen operations, making it the most comprehensive kitchen manipulation dataset for training general-purpose cooking robots.
Key specifications
- Trajectories: 55,000
- Duration: 600 hours
- Tasks: 316 distinct kitchen manipulation tasks
- Robot: Franka Panda on Omron mobile base
- Format: HDF5, Parquet, MP4
- License: CC BY 4.0 — commercial use permitted
How it compares
RoboCasa covers 100 kitchen tasks in simulation with 100K demonstrations. PhysicalAI Kitchen Demos covers 316 kitchen tasks in real hardware with 55K trajectories. Real data has higher fidelity but is harder to scale. Together they provide the most comprehensive kitchen robot training resource available — simulation breadth + real-world fidelity.
Limitations and access notes
Single robot platform (Franka Panda on Omron base) — policies may require adaptation for other kitchen robot hardware. CC BY 4.0 permits commercial use with attribution.
Linked professions
Frequently asked questions
How does PhysicalAI Kitchen Demos compare to RoboCasa?
RoboCasa is simulation-based with 100 kitchen tasks and 100K demonstrations. PhysicalAI Kitchen Demos is real-robot teleoperation with 316 kitchen tasks and 55K demonstrations. Real data offers higher fidelity; simulation offers broader coverage. They are complementary — researchers typically use simulation for pre-training and real data for fine-tuning.
Why does the robot use a mobile base in this dataset?
A mobile base (Omron) allows the robot to navigate around the kitchen rather than operating from a fixed position. Real kitchen tasks require moving between counter, stove, refrigerator, and sink — a capability stationary arm datasets cannot train. The mobile base data enables whole-body kitchen navigation policies.
Can PhysicalAI Kitchen Demos be used commercially?
Yes. CC BY 4.0 licensed, permitting commercial use with attribution to NVIDIA.
What are the 316 kitchen tasks?
The 316 tasks span the full kitchen operation spectrum: food preparation (chopping, mixing, measuring), cooking (stovetop, oven, microwave operation), cleaning (wiping, washing, tidying), appliance operation, storage and retrieval, and multi-step recipe execution requiring task sequencing across multiple kitchen stations.
How do I access PhysicalAI Kitchen Demos?
Available on Hugging Face at huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Manipulation-Kitchen-Demos. No registration required.