Robbyant Releases LingBot-VLA 2.0 Open Source 6B Cross Entity Robot Vision Language Action Model

Robbyant has released LingBot-VLA 2.0, an open source 6B parameter vision language action model built to control multiple robot platforms through one unified representation.

The model uses a MoE action expert design on top of a Qwen3-VL-4B-Instruct backbone. It encodes robot states and actions as 55 dimensional normalized vectors. This approach allows the same weights to drive different robot bodies without separate fine tuning for each hardware type.

Training data reached about 60,000 hours. Of that total, 50,000 hours came from robot trajectories and 10,000 hours came from first person human video. The data set covers 20 distinct robot configurations.

On the GM-100 dual arm benchmark the new model posted higher success rates than both the earlier LingBot-VLA version and the π0.5 baseline across several tested platforms.

Model weights, training code and the technical report are now available under the Apache 2.0 license on public repositories.

The release follows a pattern seen in other recent open source robot foundation models where teams publish both the weights and the data processing pipeline together.

Unified action space reduces per robot retraining

The 55 dimensional vector format replaces separate action heads that normally must be rewritten when hardware changes. A single set of expert modules now handles joint positions, gripper commands and base velocities for arms, mobile bases and dual arm systems.

Robbyant states that inference runs at 30 Hz on an RTX 4090 class GPU when the model is quantized to 8 bit. Latency measurements were taken on a standard x86 workstation with no custom accelerators.

Developers can load the checkpoint and swap only the robot description file to move from one platform to another. The company published example configuration files for five common research robots.

Early users report that zero shot transfer works best when the target robot stays within the kinematic range seen during training. Out of distribution morphologies still require small adapter layers.

Data volume and diversity drive the reported gains

The 60,000 hour corpus is roughly three times larger than the data used for the first LingBot-VLA release. The added first person human video supplies dense visual motion cues that are difficult to collect from robot sensors alone.

Robbyant filtered the human video through pose estimation and retargeted the resulting trajectories to robot joint spaces. This step created additional synthetic robot demonstrations without new hardware collection.

Benchmark tables show the largest lift on tasks that involve long horizon planning and contact rich manipulation. Short pick and place motions improved by smaller margins.

Independent replication of the GM-100 numbers has not yet appeared in peer reviewed venues. The published technical report contains the evaluation protocol and seed values so others can rerun the tests.

Open weights create pressure on closed robot foundation labs

Several commercial robot AI teams have kept model weights private while selling access through cloud APIs. The new release places direct performance numbers in the public domain for the first time on the GM-100 suite.

Teams that previously cited data scarcity as the main barrier now face a public baseline trained on 60,000 hours. Any new proprietary model will need to demonstrate clear gains over the open checkpoint to justify closed distribution.

Hardware vendors who sell arms and mobile bases can now point customers to a single model instead of maintaining multiple fine tuned weights. This reduces their internal engineering load for software updates.

Limits remain on long horizon reliability and safety

The technical report lists several failure modes observed during evaluation. These include occasional drift in base velocity commands after more than 40 steps and reduced precision when gripper contact forces exceed the range seen in training data.

No formal safety certification or formal verification of the model outputs is included in the current release. Users must still implement their own collision checks and emergency stop logic.

Robbyant acknowledges that sim to real gaps persist on tasks that involve deformable objects or variable lighting. The report recommends domain randomization during fine tuning for such cases.

Next signals to watch in the coming months

Three concrete milestones will indicate whether the model gains traction beyond research labs.

First, watch for independent replication studies posted on arXiv that rerun the GM-100 evaluation with different random seeds. Matching or exceeding the published scores on new hardware would strengthen the claim of broad transfer.

Second, monitor whether any of the major robot arm manufacturers integrate the LingBot-VLA 2.0 checkpoint into their official SDK examples. Official support would indicate commercial interest in the unified action space approach.

Third, track the volume of downstream fine tuned variants uploaded to public hubs. A steady increase in community fine tunes over the next quarter would show that the base model serves as a useful starting point rather than a one off release.

Developers working on multi robot fleets or comparative studies of open versus closed VLA models now have a concrete public reference point. The coming months will show whether that reference changes deployment patterns or remains limited to academic benchmarks.