Short instructions for controlling robots with Verne V0.5

The open weights model combines video learning with robot commands. Its demonstrations pair learned skills with a separate planning layer.

Two robots share a logistics task: the first retrieves a bin from a shelf, the second scans and sorts its contents, and the first returns the empty bin. In this Verne V0.5 demonstration, each machine executes short instructions. A separate reasoning layer, identified as ChatGPT Astra in the launch posts, tracks their progress and assigns the next steps.

The robot control model receives camera images, the robot’s state, and an instruction in English. It maintains a memory of the episode and produces upcoming movements. This division of responsibilities also supports assembling a makeup kit: the planning layer checks which products are already inside, requests the remaining items, and can direct the robot to retrieve a dropped item before continuing. Released weights with commercial conditions

The base model weights are available upon request, following review by the team and acceptance of the license. Two additional versions are fine-tuned for the LIBERO and RoboTwin 2.0 simulation environments. The release includes the model code and a minimal inference package, but no training or fine-tuning code. The provider offers a platform for adapting the model using a user’s demonstrations.

The Verne V0.5 Model License permits noncommercial research regardless of revenue. Royalty-free commercial use is allowed for organizations with annual revenue of up to $5 million, including affiliated entities. Above that threshold, a separate commercial agreement is required. The open weights designation therefore does not mean the release carries an open-source license without commercial restrictions. Video prediction is part of training

The architecture starts from the Wan2.2 video model, adapted to robotic observations, then combines a video transformer and an action transformer trained together. The first learns how a scene changes; the second connects those changes to the robot’s movements. The reported pretraining dataset contains more than 1.7 million demonstrations from vendors, public datasets, and internal collection.

During execution, no future video is generated. Using three camera views and its observation history, the network produces 32 motion targets at 30 Hz, covering approximately 1.07 seconds. These describe the positions and orientations of the tools at the ends of the arms, gripper openings, and mobile base movement. An execution layer then converts those targets into trajectories specific to the machine.

The reported 8.2 billion parameters refer to the main weights file; the text encoder and video compression component load separately. The inference repository specifies an NVIDIA GPU with at least 24 GB of memory. The base model still needs fine-tuning for the robot, its cameras, and the intended tasks. Reported gains come from simulated environments

In the LIBERO evaluation card, average success reaches 99.3% across 2,000 episodes, compared with 96.85% for π0.5 in the published comparisons. The Long suite, which covers multistep tasks, reaches 99.2%. The claimed lead applies to open weights models developed by U.S. companies: the provider’s table lists ABot-M0.5 with a slightly higher overall average of 99.4%.

Across the 50 two-arm tasks in RoboTwin 2.0, reported success is 81% in clean scenes and 73.4% when clutter, lighting, textures, or table heights change. The comparison score for π0.5 falls to 46% in those randomized scenes, a gap of 27.4 percentage points. Other models rank ahead of V0.5 in the same table. Fine-tuning uses only clean scenes, with 50 demonstrations per task.

These rates come from simulations and do not measure reliability on a production floor. The physical robot videos show other tasks, including loading a cart after less than two hours of fine-tuning data, according to the team. That duration refers to the demonstrations collected, not the total training time. The controller affects movement precision

Continuous motion also depends on the software connecting predictions to motors. While the network prepares a new sequence, the robot continues executing committed movements; the next sequence preserves those targets to avoid a discontinuity. A separate controller runs at 500 Hz.

An internal comparison involving LEGO disassembly uses the same V0.5 policy with two controllers, across 20 trials each. Reported median tracking error ranges from 9 to 26 millimeters with the original YAM controller, compared with 2.2 to 3.5 millimeters with Verne’s controller, depending on speed.