Robot Training Datasets
Real-robot demonstrations, human video, and supply chain intelligence data across manipulation, locomotion, and dexterous tasks.
41 datasets · 23 real-robot · 7 human video · 32 commercial use OK
AgiBot
AgiBot World 2026
2026
Axis Robotics
Axis Sim Dataset V1
2026
Build AI
Egocentric-100K — Manual Labor Egocentric Video Dataset
2026
Build AI
Egocentric-1M — The Internet for Physical AI
2026
Galaxea Robotics
Galaxea Open-World Dataset — Mobile Manipulation
2026
NVIDIA
GR1 Robot Teleoperation Dataset
2026
Robotic Systems Lab
GrandTour — Legged Robot Dataset in the Wild
2026
Humanity's Last Machine
Humanity's Last Machine — Humanoid Robot Supplier Database
2026
LMMs Lab
OmniAction — Multimodal Proactive Robot Manipulation Dataset
2026
PalmDex Research Team
PalmDex — Embodiment-Agnostic Tactile Manipulation Dataset
2026
NVIDIA
PhysicalAI Robotics GR00T X-Embodiment Sim
2026
NVIDIA
PhysicalAI Robotics Kitchen Manipulation Demos
2026
Stanford University
RoboReward — Vision-Language Reward Dataset for Robotics
2026
Intuitive Robots Research Team
SPARC (Spatial Annotations from Robot Demonstrations at Scale)
2026
Xperience Research Consortium
Xperience-10M — Large-Scale Egocentric Multimodal Human Experience Dataset
2026
AgiBot
AgiBot-Alpha — AgiBot World Dataset
2025
AI Robotic Association (AIRoA)
AIRoA Mobile Manipulator Data Engine
2025
Stanford University REAL Lab
DexUMI — Dexterous Universal Manipulation Interface
2025
Carnegie Mellon University
DexWild — In-the-Wild Dexterous Manipulation Dataset
2025
GaTech RL2
EgoVerse — Egocentric Robot Training Dataset: 8.5M Frames, 74 Countries
2025
Rniu Lab (chrisyrniu)
Human2LocoMan — Loco-Manipulation Dataset
2025
Hugging Face
LeRobot Community Datasets v3
2025
DAVIAN Robotics
PHUMA — Physically-Grounded Humanoid Locomotion Dataset
2025
NVIDIA
PhysicalAI Robotics GR00T Teleop G1
2025
RealMan Robotics
RealSource-World — Dual-Arm Humanoid Manipulation Dataset
2025
Claru AI
Claru — Training Data for Physical AI
2024
Stanford University
DROID — A Large-Scale In-The-Wild Robot Manipulation Dataset
2024
Fourier Intelligence
Fourier ActionNet — Humanoid Manipulation Dataset
2024
Julian Saks
Humanoid Atlas — Global OEM & Supplier Supply Chain Database
2024
UT Austin
RoboCasa — Large-Scale Simulation for Everyday Household Tasks
2024
X-Humanoid
RoboMIND — Multi-Embodiment Intelligence Normative Data
2024
Stanford University
ALOHA — Action Chunking with Transformers
2023
UC Berkeley RAIL
BridgeData V2
2023
UT Austin
LIBERO — Lifelong Learning for Robot Manipulation
2023
Google DeepMind
Open X-Embodiment — Robotic Learning Datasets
2023
Hugging Face
PushT — Diffusion Policy Reference Dataset
2023
UT Austin RAIL Lab
RoboSet — A Diverse Multi-Task Robot Learning Dataset
2023
Meta AI Research
Ego4D — Around the World in 3,025 Hours of Egocentric Video
2022
University of Freiburg
CALVIN — Composing Actions from Language and Vision
2021
Stanford University
RoboMimic — A Study of Imitation Learning from Observations
2021
UC Berkeley
Meta-World — Multi-Task and Meta-Reinforcement Learning
2019
About Robot Training Datasets
Robot training datasets are structured collections of demonstrations, sensor recordings, and task annotations used to train AI models for autonomous robot control. They power imitation learning, reinforcement learning, and vision-language-action model fine-tuning — the core techniques behind modern humanoid and manipulation robots. Open datasets such as EgoVerse (57,761 episodes across household manipulation tasks, CC BY-SA 4.0), Open X-Embodiment (over one million cross-embodiment demonstrations, Apache 2.0), and DROID (76,000 real-world manipulation episodes, Apache 2.0) provide foundations for general-purpose robot policies. Humanoid-specific datasets including PhysicalAI GR00T Teleop G1 (Unitree G1 teleoperation, CC BY 4.0) and PHUMA (locomotion for Unitree G1 and H1_2, Apache 2.0) enable training directly on production humanoid platforms. Dexterous manipulation datasets such as DexWild (CMU, 93 environments, MIT) and DexUMI (Stanford REAL Lab, Inspire Hand, MIT) address fine motor control requiring six-degree-of-freedom hand dexterity.
Dataset Types
Real Robot
Demonstrations collected directly on physical robot hardware. Captures sensor data, motor commands, and visual observations during actual task execution. Highest fidelity — what the robot actually experienced.
Human Video
Human demonstrations recorded and retargeted to robot kinematic structures. Enables diverse and natural motion data at scale without requiring robot hardware for collection.
Synthetic
Data generated in simulation environments, enabling large-scale collection across edge cases and scenarios impractical to collect physically.
Supply Chain Intelligence
Structured datasets documenting hardware supplier relationships, component specifications, and OEM networks in the robotics industry.
Frequently Asked Questions
What open-source robot training datasets are available in 2025 and 2026?
Geppetto indexes 41 robot training datasets as of 2026. Major open datasets include Axis Sim Dataset V1 (Axis Robotics, open-source), RoboReward — Vision-Language Reward Dataset for Robotics (Stanford University, CC BY 4.0), PhysicalAI Robotics GR00T X-Embodiment Sim (NVIDIA, CC BY 4.0), RoboMIND — Multi-Embodiment Intelligence Normative Data (X-Humanoid, Apache 2.0), LeRobot Community Datasets v3 (Hugging Face, Apache 2.0), PhysicalAI Robotics Kitchen Manipulation Demos (NVIDIA, CC BY 4.0), GR1 Robot Teleoperation Dataset (NVIDIA, MIT), PalmDex — Embodiment-Agnostic Tactile Manipulation Dataset (PalmDex Research Team, CC BY-SA 4.0).
Which robot training datasets allow commercial use?
Robot training datasets with commercial-use licences in the Geppetto directory include Axis Sim Dataset V1 (open-source), Egocentric-1M — The Internet for Physical AI (Apache 2.0), RoboReward — Vision-Language Reward Dataset for Robotics (CC BY 4.0), PhysicalAI Robotics GR00T X-Embodiment Sim (CC BY 4.0), Xperience-10M — Large-Scale Egocentric Multimodal Human Experience Dataset (Apache 2.0), Egocentric-100K — Manual Labor Egocentric Video Dataset (Apache 2.0), RoboMIND — Multi-Embodiment Intelligence Normative Data (Apache 2.0), LeRobot Community Datasets v3 (Apache 2.0), PhysicalAI Robotics Kitchen Manipulation Demos (CC BY 4.0), GR1 Robot Teleoperation Dataset (MIT). Geppetto indexes 32 commercially licensed datasets in total.
What is a robot training dataset?
A robot training dataset is a structured collection of demonstrations, sensor observations, and task labels recorded during robot or human task performance. These datasets are used to train machine learning models for robot control through imitation learning, behaviour cloning, and vision-language-action model fine-tuning. They typically include camera images, depth data, proprioceptive readings, motor commands, and natural language task descriptions. Geppetto indexes 41 robot training datasets across real-robot, human video, synthetic, and supply chain intelligence categories.
What robot training datasets exist for humanoid robots?
Humanoid robot training datasets in the Geppetto directory include AgiBot World 2026 (AgiBot, 2026), PhysicalAI Robotics GR00T X-Embodiment Sim (NVIDIA, 2026), OmniAction — Multimodal Proactive Robot Manipulation Dataset (LMMs Lab, 2026), RoboMIND — Multi-Embodiment Intelligence Normative Data (X-Humanoid, 2024), GR1 Robot Teleoperation Dataset (NVIDIA, 2026), Fourier ActionNet — Humanoid Manipulation Dataset (Fourier Intelligence, 2024), AgiBot-Alpha — AgiBot World Dataset (AgiBot, 2025), RealSource-World — Dual-Arm Humanoid Manipulation Dataset (RealMan Robotics, 2025), PHUMA — Physically-Grounded Humanoid Locomotion Dataset (DAVIAN Robotics, 2025), PhysicalAI Robotics GR00T Teleop G1 (NVIDIA, 2025).
What are the largest robot training datasets by episode count?
The largest robot training datasets by episode count include Xperience-10M — Large-Scale Egocentric Multimodal Human Experience Dataset (10,000,000 episodes, Apache 2.0), Open X-Embodiment — Robotic Learning Datasets (1,000,000 episodes, Apache 2.0), Meta-World — Multi-Task and Meta-Reinforcement Learning (500,000 episodes, MIT), LIBERO — Lifelong Learning for Robot Manipulation (130,000 episodes, MIT), RoboMIND — Multi-Embodiment Intelligence Normative Data (107,000 episodes, Apache 2.0). The largest by recording hours include Egocentric-1M — The Internet for Physical AI (1,000,000 hours), Egocentric-100K — Manual Labor Egocentric Video Dataset (100,405 hours), AIRoA Mobile Manipulator Data Engine (10,000 hours).
What robot training datasets are available for dexterous manipulation?
Dexterous manipulation robot training datasets include RoboMIND — Multi-Embodiment Intelligence Normative Data (X-Humanoid, Apache 2.0), LeRobot Community Datasets v3 (Hugging Face, Apache 2.0), GR1 Robot Teleoperation Dataset (NVIDIA, MIT), PalmDex — Embodiment-Agnostic Tactile Manipulation Dataset (PalmDex Research Team, CC BY-SA 4.0), RoboMimic — A Study of Imitation Learning from Observations (Stanford University, MIT), Fourier ActionNet — Humanoid Manipulation Dataset (Fourier Intelligence, Fourier Research License — non-commercial research use), ALOHA — Action Chunking with Transformers (Stanford University, MIT), DexWild — In-the-Wild Dexterous Manipulation Dataset (Carnegie Mellon University, MIT), DexUMI — Dexterous Universal Manipulation Interface (Stanford University REAL Lab, MIT — confirm in dataset directory before activation).