HOURS ≠ TRAJECTORIES ≠ SIMULATION

Physical AI Training Data: What the Counts Measure

Compare six reported robot-learning data budgets by training stage, unit, included subsets and access limits. Counts are reported by the authors, not independently recounted.

6 retained recordsSources checked September 29, 2026Primary-source fields preserved per record

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

LingBot-VLA 2.0 pretraining mix

Official repository

Dataset / version: LingBot-VLA 2.0 pretraining mix
Training stage: Pretraining mixture
Author-reported scale: ~50,000 robot hours + ~10,000 egocentric human-video hours
Inputs / measurements: Vision, language, robot state/action; human egocentric video in pretraining mix
Why this is not directly comparable: Human video and robot-action data are distinct components. The 60,000-hour total does not mean 60,000 robot-control hours.
Available artifacts / access: Repository links code and model artifacts. It does not present the complete proprietary pretraining corpus as a single downloadable dataset.
Source checked: 2026-09-29
Source: LingBot-VLA 2.0

OpenVLA pretraining data

Official project

Dataset / version: OpenVLA pretraining data
Training stage: Pretraining mixture
Author-reported scale: 970,000 robot episodes from Open X-Embodiment
Inputs / measurements: RGB, language and robot actions/state depending on source dataset
Why this is not directly comparable: Episode duration varies and no comparable hour total is given. Some underlying datasets also appear in other mixtures.
Available artifacts / access: Model/code are released; underlying Open X-Embodiment datasets have separate access and licensing terms.
Source checked: 2026-09-29
Source: OpenVLA project

RDT-1B multi-robot pretraining collection

Official project

Dataset / version: RDT-1B multi-robot pretraining collection
Training stage: Pretraining and fine-tuning
Author-reported scale: 1M+ pretraining episodes from 46 datasets; separately, 6K+ ALOHA fine-tuning episodes
Inputs / measurements: Language, up to three RGB views, robot actions/state
Why this is not directly comparable: The 46-dataset mix and the authors' ALOHA collection describe separate stages. Shared dataset families prevent adding this budget to OpenVLA's.
Available artifacts / access: Code, weights and the authors' ALOHA fine-tuning dataset are linked. The model card's MIT statement is not a blanket licence for all 46 source datasets.
Source checked: 2026-09-29
Source: RDT-1B project RDT-1B model card ALOHA fine-tuning dataset

RoboMIND 2.0

Research paper

Dataset / version: RoboMIND 2.0
Training stage: Dataset release
Author-reported scale: 310K+ real-world trajectories; an additional 20K simulated trajectories
Inputs / measurements: Multimodal robot data; includes 12K tactile-enhanced and 20K mobile-manipulation trajectories
Why this is not directly comparable: The 12K tactile and 20K mobile-manipulation trajectories are subsets of the real-world data, not extra totals. Six embodiments and 739 tasks describe coverage.
Available artifacts / access: The paper describes the dataset. Complete file availability and dataset licence were not established by this audit.
Source checked: 2026-09-29
Source: RoboMIND 2.0 paper, v3

TurboVLA AgileX Piper task demonstrations

Research result

Dataset / version: TurboVLA AgileX Piper task demonstrations
Training stage: Real-robot fine-tuning
Author-reported scale: 65 teleoperated demonstrations per task × 4 tasks = 260
Inputs / measurements: Wrist-view RGB-D, third-view RGB-D, language, robot action/state
Why this is not directly comparable: This is not the whole training budget: the paper pretrains on LIBERO before Piper fine-tuning. Two RGB-D cameras are documented; no footage was used for this count.
Available artifacts / access: The v3 paper documents the real-robot protocol. This audit did not establish access to the full 260-demonstration corpus.
Source checked: 2026-09-29
Source: TurboVLA paper, v3 (27 September 2026)TurboVLA project (different results version)

Isaac GR00T N1.7 training mixture

Official repository / paper

Dataset / version: Isaac GR00T N1.7 training mixture
Training stage: Pretraining mixture
Author-reported scale: 20,000 human-video hours reported for N1.7; no combined robot-hour total used here
Inputs / measurements: Images, language, robot state/action, synthetic trajectories
Why this is not directly comparable: This row now identifies N1.7 explicitly. Its human-video quantity is not a total for all GR00T N1 releases and is not robot demonstration time.
Available artifacts / access: N1.7 code and model release are documented. The repository does not establish a single downloadable package for the entire mixed training corpus.
Source checked: 2026-09-29
Source: NVIDIA Isaac GR00T

Dataset / version	Training stage	Author-reported scale	Inputs / measurements	Why this is not directly comparable	Available artifacts / access	Source checked	Source
LingBot-VLA 2.0 pretraining mix	Pretraining mixture	~50,000 robot hours + ~10,000 egocentric human-video hours	Vision, language, robot state/action; human egocentric video in pretraining mix	Human video and robot-action data are distinct components. The 60,000-hour total does not mean 60,000 robot-control hours.	Repository links code and model artifacts. It does not present the complete proprietary pretraining corpus as a single downloadable dataset.	2026-09-29	LingBot-VLA 2.0
OpenVLA pretraining data	Pretraining mixture	970,000 robot episodes from Open X-Embodiment	RGB, language and robot actions/state depending on source dataset	Episode duration varies and no comparable hour total is given. Some underlying datasets also appear in other mixtures.	Model/code are released; underlying Open X-Embodiment datasets have separate access and licensing terms.	2026-09-29	OpenVLA project
RDT-1B multi-robot pretraining collection	Pretraining and fine-tuning	1M+ pretraining episodes from 46 datasets; separately, 6K+ ALOHA fine-tuning episodes	Language, up to three RGB views, robot actions/state	The 46-dataset mix and the authors' ALOHA collection describe separate stages. Shared dataset families prevent adding this budget to OpenVLA's.	Code, weights and the authors' ALOHA fine-tuning dataset are linked. The model card's MIT statement is not a blanket licence for all 46 source datasets.	2026-09-29	RDT-1B project RDT-1B model card ALOHA fine-tuning dataset
RoboMIND 2.0	Dataset release	310K+ real-world trajectories; an additional 20K simulated trajectories	Multimodal robot data; includes 12K tactile-enhanced and 20K mobile-manipulation trajectories	The 12K tactile and 20K mobile-manipulation trajectories are subsets of the real-world data, not extra totals. Six embodiments and 739 tasks describe coverage.	The paper describes the dataset. Complete file availability and dataset licence were not established by this audit.	2026-09-29	RoboMIND 2.0 paper, v3
TurboVLA AgileX Piper task demonstrations	Real-robot fine-tuning	65 teleoperated demonstrations per task × 4 tasks = 260	Wrist-view RGB-D, third-view RGB-D, language, robot action/state	This is not the whole training budget: the paper pretrains on LIBERO before Piper fine-tuning. Two RGB-D cameras are documented; no footage was used for this count.	The v3 paper documents the real-robot protocol. This audit did not establish access to the full 260-demonstration corpus.	2026-09-29	TurboVLA paper, v3 (27 September 2026)TurboVLA project (different results version)
Isaac GR00T N1.7 training mixture	Pretraining mixture	20,000 human-video hours reported for N1.7; no combined robot-hour total used here	Images, language, robot state/action, synthetic trajectories	This row now identifies N1.7 explicitly. Its human-video quantity is not a total for all GR00T N1 releases and is not robot demonstration time.	N1.7 code and model release are documented. The repository does not establish a single downloadable package for the entire mixed training corpus.	2026-09-29	NVIDIA Isaac GR00T

Methodology

These figures answer different questions. LingBot and OpenVLA describe pretraining; TurboVLA's 260 demonstrations describe real-robot fine-tuning after LIBERO pretraining. RoboMIND counts a dataset, not every sample consumed by a model. RDT and OpenVLA draw on overlapping dataset families, so their totals must not be added. Hours are not converted to episodes. A released model or code repository does not establish that its complete training corpus can be downloaded.

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.