LingBot-VLA 2.0

LingBot-VLA 2.0 is a 6-billion-parameter vision-language-action model developed by Robbyant, the embodied-intelligence unit of Ant Group. It maps camera observations and natural-language instructions to robot actions across several types of manipulators and humanoid systems. Robbyant released a technical report, source code, and pretrained weights in July 2026.^[1][2][3]

The model extends the first LingBot-VLA from dual-arm manipulation toward cross-embodiment control that can include arms, grippers, dexterous hands, waist, head, and mobile-base movement. Its central research claim is that one policy can share general manipulation knowledge while retaining specialists for different bodies and tasks.^[1]

Scroll sideways for more →

Architecture

LingBot-VLA 2.0 uses Qwen3-VL-4B-Instruct as its vision-language backbone and adds an action expert that predicts chunks of continuous robot controls.^[1] The action expert replaces ordinary dense feed-forward layers with sparse mixture-of-experts layers. Robbyant describes fine-grained expert segments plus a shared expert, allowing the model to route tokens toward specialized computation without activating every parameter for every input.^[1][2]

Robot states and actions are converted to a canonical 55-dimensional representation. It reserves dimensions for left and right arms, end effectors or grippers, dexterous hands, waist, head, and mobile base. Dimensions absent from a particular robot are masked or padded. The paper covers 20 configurations from 17 manufacturers, spanning single-arm, dual-arm, wheeled, and humanoid systems.^[1] This is a common interface for training, not proof that an unconfigured robot will work without calibration, action mapping, normalization statistics, and task-specific post-training.

The second architectural change is dual-query distillation. The model appends one query for the current visual scene and one for a future scene. During training, these queries learn from LingBot-Depth for geometric information and DINO-Video for temporal information.^[1][2] Robbyant frames future-image prediction as a proxy for learning action consequences. The deployed policy does not require those teacher models to generate every action, but training and reproduction require their checkpoints and configuration.^[3]

Training data

The authors report roughly 60,000 hours of curated pretraining material: about 50,000 hours of robot trajectories across 20 configurations and 10,000 hours of first-person human manipulation video.^[1] This triples the roughly 20,000 robot hours and nine configurations reported for the original LingBot-VLA.^[4]

The paper describes filtering for motion quality and image quality, removing episodes with abnormal motion, long static periods, blur, occlusion, missing frames, or synchronization problems.^[1] The released repository contains training code and configuration support, but the full 60,000-hour corpus is not distributed with the model. Consequently, the checkpoint is public while exact pretraining is not independently reproducible from the released artifacts alone.

Evaluation

Robbyant evaluates generalization on GM-100, a 100-task real-robot benchmark used in the earlier LingBot work, and evaluates longer mobile-manipulation tasks on two wheeled platforms.^[1] Comparisons include the original LingBot-VLA, OpenPI-derived baselines, pi0.5, and NVIDIA GR00T N1.7. The paper reports both task-progress and binary success measures because a robot may complete several subtasks but fail the whole sequence.^[1]

These are author-run evaluations, not independent leaderboard measurements. Hardware embodiments, post-training demonstrations, task resets, action mappings, and stopping criteria can materially affect real-robot results. The paper's cross-embodiment findings show that one trained representation can be adapted across the evaluated platforms; they do not establish universal control of arbitrary robots. The public repository also had open requests for missing configuration examples and minimum deployment GPU requirements after release, illustrating remaining reproduction work.^[5]

Released artifacts and licensing

Robbyant's GitHub repository contains training, post-training, evaluation, and deployment code. Its project metadata and top-level license identify Apache License 2.0.^[3] The official Hugging Face model card also labels the 6B checkpoint Apache-2.0 and reports six billion parameters stored in safetensors.^[2] License statements are artifact-specific: dependencies and the Qwen3-VL backbone retain their own terms, so the repository's license does not replace every upstream license.

The repository also lists a RoboTwin 2.0 post-training checkpoint released on July 25, 2026.^[3] That checkpoint is distinct from the general pretrained robbyant/lingbot-vla-v2-6b model. Users need the matching robot configuration, camera names, joint normalization statistics, and processor assets when fine-tuning or deploying.^[3]

OpenRoboto integration

Robbyant said on August 31, 2026 that OpenRoboto had brought LingBot-VLA 2.0 to its platform as a base model for community post-training.^[6] That statement verifies platform integration at a high level. It does not identify a public OpenRoboto checkpoint, evaluation protocol, task result, or reproduction of Robbyant's GM-100 and cross-embodiment claims. An OpenRoboto Hugging Face URL surfaced during research, but unauthenticated page and API requests returned HTTP 401 on September 2, 2026. Its model card and files therefore could not be independently inspected, and the repository's public, gated, or private status could not be distinguished from the unauthenticated response alone.

RealMan partnership

RealMan Robotics described itself as a hardware partner for the release and said its RS-02 data-collection robot fed data into the pretraining pipeline.^[7] The LingBot paper does not quantify data by supplier or name RealMan in its abstract, so RealMan's share of the 50,000 robot hours is unknown. A separate RealMan X post showing RealBOT material-handling training is not evidence for the LingBot partnership and is not used for that claim.

Limitations

LingBot-VLA 2.0 is an open-weight research system rather than a turnkey universal robot controller. The full pretraining corpus is unavailable, deployment requirements depend on the embodiment, and the headline evaluations were conducted by the developers. Safe physical deployment also requires controls outside the learned policy, including collision limits, emergency stops, workspace constraints, and human supervision. The available sources do not establish safety certification or production reliability for the model itself.

See also

• Robot learning
• Embodied AI
• Robbyant R1
• OpenVLA

References

1. ^Wu, Wei, et al. "From Foundation to Application: Improving VLA Models in Practice." arXiv:2607.06403, July 2026. arxiv.org/...2607.06403
2. ^Robbyant. "lingbot-vla-v2-6b" model card. Hugging Face. Accessed September 2, 2026. huggingface.co/...lingbot-vla-v2-6b
3. ^Robbyant. "lingbot-vla-v2: From Foundation to Application." GitHub. Accessed September 2, 2026. github.com/...lingbot-vla-v2
4. ^Robbyant Team, et al. "A Pragmatic VLA Foundation Model." arXiv:2601.18692, January 2026. arxiv.org/...2601.18692
5. ^Robbyant. LingBot-VLA 2.0 repository issues. GitHub. Accessed September 2, 2026. github.com/...issues
6. ^Robbyant (@robbyant_brain). "Glad to see @openroboto bring LingBot-VLA 2.0 to its platform as a base model for community post-training." X, August 31, 2026. x.com/...2094431739347410967
7. ^RealMan Robotics. "RealMan Joins Robbyant's LingBot-VLA 2.0 Open-Source Push." LinkedIn, 2026. linkedin.com/...-activity-7481723983902466048-ly9X