LingBot-VLA 2.0 Open Source Release and Robot Coverage

The public repository contains installation instructions, model download commands, training configuration, open-loop evaluation, RoboTwin deployment and a real-robot policy server. Robbyant links a 6B pretrained native-depth checkpoint from the repository.

The Apache-2.0 license makes the codebase usable for research and development under the license terms. Public code does not mean the full training dataset, every internal tool or every evaluation environment has been released.

Robbyant’s paper defines the pretraining corpus as 50,000 hours of cleaned robot trajectories across 20 configurations plus 10,000 hours of filtered egocentric human-operation video. The human video stream is used to supply manipulation-centric motion and semantic priors after reconstruction and quality control.

This resolves the apparent difference between 50,000 and 60,000 hours. The smaller number is the retained robotic subset. The larger number is the combined pretraining corpus.

The official paper and repository describe 20 robot configurations. They include single-arm and dual-arm systems, half-humanoids, humanoids and mobile manipulators, with named platforms such as Astribot S1 and Leju KUAVO 4 Pro.

The TechniaHQRobot announcement associated with the release states 17 robot brands. The official paper does not expose a separate brand-count field, so this article preserves 17 as a source-post claim rather than deriving it from the configuration table.

LingBot-VLA 2.0 maps heterogeneous platforms into a 55-dimensional canonical state and action vector. It allocates fields for arm joints, end-effector pose, grippers, dexterous hands, waist, head and mobility signals, with padding where a body lacks a component.

This representation lets one model ingest trajectories from robots with different shapes. It does not make their dynamics identical. Post-training, normalization and configuration mapping are still required for a target platform.

The repository provides a path from model download to post-training and deployment. Real hardware still needs camera calibration, action mapping, safety limits, latency measurement, low-level control and a dataset that matches the target task.

Robbyant reports one inference call taking about 130 milliseconds on an NVIDIA GeForce RTX 4090D with 10 denoising steps for the documented deployment path. That number belongs to the stated configuration and should not be generalized to other GPUs or complete robot response time.

The public repository, license and model links prove that a usable LingBot-VLA 2.0 codebase and pretrained checkpoint have been released. The paper documents the intended data composition, unified action representation and evaluation tasks.

The accompanying demonstrations show the model operating on the authors’ selected robot platforms and tasks. They support reproducible inspection of the released model, but they do not expose the full proprietary pretraining corpus.

Open-source availability does not prove that the model will reproduce Robbyant’s results on different hardware, cameras, data quality or task definitions. Users still need to measure success, latency, interventions and safety on their own system.

The release does not establish universal cross-embodiment control, unattended production deployment or reliable operation in every home and factory. The 20-configuration pretraining coverage is broader training evidence, not a guarantee for each platform.

Researchers can inspect the architecture, load the checkpoint, prepare a LeRobot-format dataset and test post-training on a defined task. That is more useful than a video-only announcement because failures can be measured against the same code path.

The next strong evidence will come from independent reproductions that publish hardware, dataset size, trial counts, success criteria, failure modes and end-to-end latency.