HOURS ≠ TRAJECTORIES ≠ SIMULATION

Physical AI Training Data Tracker

Track robot trajectories, real-robot hours, egocentric video, simulation and multimodal training data without converting incompatible units into one number.

6 retained recordsLast verified: August 13, 2026Primary-source fields preserved per record

Collection type

6 of 6 records shownFilters are local and do not create crawlable URLs

LingBot-VLA 2.0 pretraining mix

Official repository

Dataset / program: LingBot-VLA 2.0 pretraining mix
Organization: Robbyant / Ant Group
Country: China
Models: LingBot-VLA 2.0
Reported hours: ~60,000 total reported
Trajectories: Not converted from hours
Robot hours: ~50,000 robot trajectory hours
Human/video hours: ~10,000 egocentric human video hours
Embodiments: 20 configurations
Modalities: Vision, language, robot state/action; human egocentric video in pretraining mix
Collection type: Real robot + egocentric human video
Public status: Model/code open; full proprietary pretraining corpus not presented as downloadable dataset
Source: LingBot-VLA 2.0

OpenVLA pretraining data

Official project

Dataset / program: OpenVLA pretraining data
Organization: OpenVLA collaborators
Country: United States
Models: OpenVLA 7B
Reported hours: Not converted from episodes
Trajectories: 970,000 robot episodes reported
Robot hours: Not published as one comparable hour total
Human/video hours: Not used as an equivalent robot-hour metric
Embodiments: Multiple
Modalities: RGB, language and robot actions/state depending on source dataset
Collection type: Real robot datasets
Public status: Open project/model; underlying datasets retain their own access terms
Source: OpenVLA project

RDT-1B multi-robot pretraining collection

Official project

Dataset / program: RDT-1B multi-robot pretraining collection
Organization: Tsinghua University / collaborators
Country: China
Models: RDT-1B
Reported hours: Not converted from episodes
Trajectories: 1M+ episodes plus 6K+ ALOHA fine-tuning episodes
Robot hours: Not disclosed
Human/video hours: Not disclosed
Embodiments: Multi-robot
Modalities: Language, up to three RGB views, robot actions/state
Collection type: Real robot datasets
Public status: Code, weights and data resources linked by project
Source: RDT-1B project

RoboMIND 2.0

Research paper

Dataset / program: RoboMIND 2.0
Organization: Beijing Humanoid Robot Innovation Center / collaborators
Country: China
Models: XR-1 ecosystem / MIND-2 research
Reported hours: Not converted from trajectories
Trajectories: 310K+ real-world dual-arm trajectories + 20K simulated trajectories
Robot hours: Not expressed as hours
Human/video hours: Not expressed as hours
Embodiments: 6
Modalities: Multimodal robot data; includes 12K tactile-enhanced and 20K mobile-manipulation trajectories
Collection type: Real robot + simulation
Public status: Research dataset release
Source: RoboMIND 2.0 paper

TurboVLA AgileX Piper task demonstrations

Research result

Dataset / program: TurboVLA AgileX Piper task demonstrations
Organization: HUST / Huawei
Country: China
Models: TurboVLA
Reported hours: Not disclosed; not estimated from demonstration count
Trajectories: 260 teleoperated demonstrations total across four tasks
Robot hours: Not disclosed
Human/video hours: Not disclosed
Embodiments: Single-arm
Modalities: Wrist-view RGB-D, third-view RGB-D, language, robot action/state
Collection type: Teleoperation + real robot
Public status: Experiment protocol public; raw dataset availability differs from code/checkpoints
Source: TurboVLA project

GR00T N1 family training mixture

Official repository / paper

Dataset / program: GR00T N1 family training mixture
Organization: NVIDIA
Country: United States
Models: Isaac GR00T N1 family
Reported hours: Not disclosed as one comparable aggregate
Trajectories: Version-dependent
Robot hours: Real robot demonstrations included; aggregate not normalized here
Human/video hours: Human video included; quantity not normalized here
Embodiments: Multiple
Modalities: Images, language, robot state/action, synthetic trajectories
Collection type: Human video + simulation + generated trajectories + real robot
Public status: Model/code open for current N1.7 release; full training corpus is not one downloadable dataset
Source: NVIDIA Isaac GR00T

Dataset / program	Organization	Country	Models	Reported hours	Trajectories	Robot hours	Human/video hours	Embodiments	Modalities	Collection type	Public status	Source
LingBot-VLA 2.0 pretraining mix	Robbyant / Ant Group	China	LingBot-VLA 2.0	~60,000 total reported	Not converted from hours	~50,000 robot trajectory hours	~10,000 egocentric human video hours	20 configurations	Vision, language, robot state/action; human egocentric video in pretraining mix	Real robot + egocentric human video	Model/code open; full proprietary pretraining corpus not presented as downloadable dataset	LingBot-VLA 2.0
OpenVLA pretraining data	OpenVLA collaborators	United States	OpenVLA 7B	Not converted from episodes	970,000 robot episodes reported	Not published as one comparable hour total	Not used as an equivalent robot-hour metric	Multiple	RGB, language and robot actions/state depending on source dataset	Real robot datasets	Open project/model; underlying datasets retain their own access terms	OpenVLA project
RDT-1B multi-robot pretraining collection	Tsinghua University / collaborators	China	RDT-1B	Not converted from episodes	1M+ episodes plus 6K+ ALOHA fine-tuning episodes	Not disclosed	Not disclosed	Multi-robot	Language, up to three RGB views, robot actions/state	Real robot datasets	Code, weights and data resources linked by project	RDT-1B project
RoboMIND 2.0	Beijing Humanoid Robot Innovation Center / collaborators	China	XR-1 ecosystem / MIND-2 research	Not converted from trajectories	310K+ real-world dual-arm trajectories + 20K simulated trajectories	Not expressed as hours	Not expressed as hours	6	Multimodal robot data; includes 12K tactile-enhanced and 20K mobile-manipulation trajectories	Real robot + simulation	Research dataset release	RoboMIND 2.0 paper
TurboVLA AgileX Piper task demonstrations	HUST / Huawei	China	TurboVLA	Not disclosed; not estimated from demonstration count	260 teleoperated demonstrations total across four tasks	Not disclosed	Not disclosed	Single-arm	Wrist-view RGB-D, third-view RGB-D, language, robot action/state	Teleoperation + real robot	Experiment protocol public; raw dataset availability differs from code/checkpoints	TurboVLA project
GR00T N1 family training mixture	NVIDIA	United States	Isaac GR00T N1 family	Not disclosed as one comparable aggregate	Version-dependent	Real robot demonstrations included; aggregate not normalized here	Human video included; quantity not normalized here	Multiple	Images, language, robot state/action, synthetic trajectories	Human video + simulation + generated trajectories + real robot	Model/code open for current N1.7 release; full training corpus is not one downloadable dataset	NVIDIA Isaac GR00T

Methodology

Human video hours, robot hours, trajectories, episodes and simulated trajectories remain separate. The database never converts trajectories into hours unless the source provides enough timing information to do so reliably.

Included

Published model, robot, benchmark, deployment or failure evidence that can be tied to a retained source.

Not inferred

Unknown autonomy, latency, trial counts, deployment scale and failure causes remain unknown when the source does not disclose them.

Update policy

Versions and historical records remain distinguishable. A new release does not silently overwrite an older model or evaluation condition.