LingBot-VLA 2.0 Pretraining Dataset
The LingBot-VLA 2.0 Pretraining Dataset, developed by Robbyant under Ant Digital Technologies, is a large-scale multimodal robotics corpus designed to advance generalist vision-language-action foundation models. It features 60,000 hours of curated interaction data, comprising 50,000 hours of cleaned real-robot trajectories and 10,000 hours of distilled egocentric human manipulation video. The dataset spans 20 distinct robot morphologies from 17 global manufacturers, including bipedal humanoids, wheeled bases, and dual-arm systems from brands like Unitree, AgileX, and AgiBot. By unifying these diverse platforms into a shared 55-dimensional canonical action space, the dataset facilitates cross-embodiment learning and whole-body control. Released under the Apache 2.0 license in July 2026, the collection supports a wide scope of tasks from industrial logistics to domestic care. It is an essential resource for researchers building scalable robotic brains that require robust morphological generalization, predictive dynamics, and high-precision spatial perception across complex, long-horizon real-world deployment scenarios.
| Year | 2026 |
|---|---|
| Total hours | 60,000 |
| Embodiments | Leju, AgiBot, Unitree, AgileX, Galaxea, Galbot, Astribot, RealMan, Franka, ARX, X-Humanoid, Fourier, MagicLab, Spirit AI, Zerith, Flexiv, Qinglong |
| Task categories | cleaning, folding, manipulation, pick-and-place, bimanual, locomotion, loco-manipulation, warehouse, long-horizon, other |
| License | Apache-2.0 |
| Access | open — commercial use permitted |
| Maintainer | Robbyant / Ant Digital Technologies |
| Origin country | CN |
Methodology
The LingBot-VLA 2.0 dataset was constructed through a rigorous 'high-quality-first' curation pipeline. Starting from a raw pool of approximately 110,000 hours (90,000 robot hours and 20,000 human hours), the developers employed an automated filtering system that assessed trajectories based on mechanical smoothness (third-order jerk), velocity Z-scores, and video-state alignment. To ensure temporal and spatial consistency, the robot data was verified by replaying recorded states against the robot's URDF model, while human egocentric videos were processed through SLAM and MANO hand-pose reconstruction to align them with a unified 55-dimensional action representation.
Collection
Data was aggregated from 20 different hardware configurations across 17 manufacturers, ensuring high diversity in kinematic chains and sensory setups. Robot platforms range from standard single-arm Franka setups to complex humanoids like the Unitree G1 and Fourier GR-2. The egocentric human component provides a semantic bridge for manipulation skills, using distilled first-person video to teach the model causal relationships between hand movements and object state changes. The dataset also includes detailed annotations generated by a 27B-parameter VLM, segmenting videos into atomic subtasks.
Use-Cases
Designed for the training of generalist robotic policies, the dataset supports applications in retail sorting, warehouse logistics, and household assistance. Its whole-body action space enables coordinated control of mobile bases, heads, waists, and dexterous hands. It specifically addresses common failure modes in laboratory-trained models, such as distribution shift and poor temporal reasoning, by providing large-scale out-of-distribution scenarios and future-prediction supervision via depth and causal video distillation teachers.
Frequently asked questions
What is the composition of the LingBot-VLA 2.0 dataset?
The dataset contains 60,000 hours of pretraining data, which is split into 50,000 hours of real-robot trajectories and 10,000 hours of egocentric human manipulation video.
Which robot platforms are supported by the unified action space?
It supports 20 configurations from 17 manufacturers, including AgiBot, Unitree, AgileX, Franka, ARX, and Fourier humanoids, mapping them to a 55-dimensional canonical vector.
How was the data filtered for quality?
Robbyant used a pipeline that filtered trajectories based on jerk, velocity/acceleration Z-scores, video-state alignment, and URDF-based replay verification to remove noisy or failed episodes.
What is the role of egocentric human video in the pretraining?
The 10,000 hours of egocentric human video are distilled to provide embodiment-free manipulation priors and causal temporal dynamics, helping the model generalize beyond specific robot hardware.
Is the dataset and model available for commercial use?
Yes, the models and the associated codebase have been released under the Apache-2.0 license, which permits commercial application.