Verne Robotics has released an 8.2 billion parameter robot model with unusually thorough model cards. Verne V0.5 is an open-weight world action model with three checkpoints on Hugging Face: a pretrained base, a LIBERO fine-tune at 99.3% average success, and a RoboTwin 2.0 fine-tune at 81.0% in clean scenes and 73.4% in randomized scenes (Verne on Hugging Face). A news summary of the launch calls it the leading US-made open-weight model on those two benchmarks. Verne's own tables tell a more detailed story, so we checked.
TL;DR: the questions people ask first
| Question | Answer (from Verne's model cards) |
|---|---|
| What is it? | An 8.2B world action model that predicts future observations and robot actions |
| Checkpoints? | v0.5-base, v0.5-libero, v0.5-robotwin |
| LIBERO result? | 99.3% average over 2,000 episodes |
| RoboTwin 2.0 result? | 81.0% clean, 73.4% randomized over 9,990 episodes |
| Is it first on the leaderboards? | No, Verne's own tables list models ahead on some rows |
| Access? | Gated, manual approval, Verne license |
| Fine-tuning code? | Not released; Verne offers a hosted fine-tuning platform |
| Physical robot use? | Base is for fine-tuning; benchmark checkpoints are sim-only |
What a world action model is
Most robot policies map camera images and an instruction straight to actions. A world action model also learns to predict how the scene will change. Verne says this helps: "Learning how a scene changes helps the model learn how to act in it."
V0.5 pairs a video transformer with an action transformer that share information through masked attention. During training both learn with flow-matching objectives, one for video and one for actions. At run time, "only the action path runs. The model does not generate future video."
Inputs are current camera frames, memory of past observations, robot state and an English instruction. Output is a chunk of 32 targets at 30 Hz: tool poses, gripper openings, base poses and progress. The model is initialized from Wan2.2-TI2V-5B, with a deeper video transformer and a new action transformer, then pretrained by Verne.
Training
Verne says it pretrained V0.5 on over 1.7 million demonstrations from four sources: third-party vendors, open-source datasets, Verne teleoperation in the robot's own embodiment, and Verne handheld grippers. It filters recording timing and motion quality, cuts bad segments and aligns instructions with behavior.
Training has three stages: robot-video pretraining, joint video and action pretraining, and task post-training with fine-tuning, human corrections and reinforcement learning. The base checkpoint completes stages one and two. The benchmark checkpoints complete stage three on their benchmarks.
The results, read carefully
LIBERO
One model covers all four LIBERO suites, with 50 episodes per task.
| Suite | Success | Episodes |
|---|---|---|
| Spatial | 98.8% | 494 of 500 |
| Object | 99.8% | 499 of 500 |
| Goal | 99.2% | 496 of 500 |
| Long | 99.2% | 496 of 500 |
| Average | 99.3% | 1,985 of 2,000 |
Verne gives a 95% interval of 98.8% to 99.7% for the average. Its comparison with published results lists ABot-M0.5 at 99.4% average and 98.4% on Long, V0.5 at 99.3% and 99.2%, then VPP2, Cosmos Policy, X-VLA, Fast-WAM, π0.5 (96.85%) and π0 (94.15%). The card says V0.5 and π0.5 tie on Spatial at 98.8% and that V0.5 leads π0.5 on Long by 6.8 points. Note that the gap among the top rows is within the stated interval, and LIBERO is close to saturated.
RoboTwin 2.0
RoboTwin 2.0 has 50 two-arm tasks. Each policy trains on 50 clean-scene demonstrations per task, then runs in clean scenes and in randomized scenes with clutter, new textures, lighting and table heights.
| Model | Clean | Randomized |
|---|---|---|
| GO-3 Nano Preview | 88.5% | 81.5% |
| PICA-WAM | 90.7% | 74.7% |
| Verne V0.5 | 81.0% | 73.4% |
| EWAM | 82.2% | 72.1% |
| ViGAR | 82.0% | 67.0% |
| π0.5 | 70.7% | 46.0% |
| π0 | 46.4% | 16.3% |
Verne says the other results are selected from the RoboTwin 2.0 leaderboard of October 30, 2026, sorted by the mean of the two columns, and that π0 is a single-task result. By that sort V0.5 is third among the rows shown. The card also says V0.5's lead over π0.5 grows from 10.3 points in clean scenes to 27.4 points in randomized scenes.
Per-task results vary widely. Hanging a mug succeeds 10% of the time clean and 16% randomized, and putting an object in a cabinet 41% and 24%, while several tasks reach 99% or 100%. The aggregate hides tasks that still fail most of the time.
Checking the headline claim
The launch summary says V0.5 "ranks as the highest performing US-made open-weight model" on the two benchmarks. What the cards support:
- Supported: V0.5 ranks near the top of the models Verne chose to show, ahead of π0.5 and π0 on both benchmarks.
- Not supported by the cards: being first overall. Four listed models, including EWAM at 82.2% and ViGAR at 82.0%, beat it on RoboTwin clean scenes, two beat it on randomized scenes, and one narrowly beats its LIBERO average. By Verne's own mean-of-two-columns sort, V0.5 is third, and EWAM is only about 0.05 points behind it.
- Not checkable here: the "US-made" and "open-weight" qualifiers. The cards do not state the nationality of the competing models, and we did not audit their licenses or weights. Verne is the one that selected the comparison rows.
None of this makes the results weak. It means the sentence "best US-made open-weight model" is a narrower claim than "best model," and it rests on a comparison set the vendor chose.
Why simulation benchmarks only go so far
LIBERO and RoboTwin 2.0 are simulation benchmarks with fixed robots, cameras and task sets. A 99.3% LIBERO score says the model has learned that benchmark's tasks very well, and the saturated top of the table means small differences matter little. RoboTwin is harder, with 50 two-arm tasks and domain randomization, which is why the spread between models is wider and why the randomized column is the more informative one.
Neither benchmark measures what an industrial buyer cares about most: reliability over thousands of cycles, recovery from unexpected states, setup time on a new cell, and safety around people. Verne's own limitations section says as much when it warns that a new robot depends on fine-tuning data in its own action space. When you evaluate V0.5 or any competing model, plan a small pilot on your robot with a fixed task list and a success count over many trials, and report failures by cause, not only the average.
How open is it?
The weights are published, but access is gated. Hugging Face shows a manual request form and says Verne reviews each request. The Verne V0.5 Model License governs use, and Verne's summary says:
- Commercial use requires contacting Verne.
- Use in countries or regions under comprehensive US sanctions, or by a person on a US restricted-party list, is not permitted.
- Weapons and unlawful surveillance are not permitted.
- Section 6 makes you responsible for physical safety measures.
The release contains weights and a minimal inference package, with no training or fine-tuning code. Verne points to its fine-tuning platform at app.vernerobotics.com. That is a commercial funnel, which is reasonable, but "open-weight" should not be read as "open source." For comparison, Ai2's ManiFlow-TB ships weights and training code under Apache 2.0.
Physical safety and limits
Verne is direct: "A machine that moves can injure people and damage property." Before running V0.5 on a robot, it says to assess risks, install an emergency stop and speed and force limits, and supervise the robot. The base card lists out-of-scope uses, including deployment without fine-tuning and use on a physical robot without a risk assessment. The benchmark checkpoints are tuned to their simulators and are not for physical robots.
Limitations the card states: performance on a new robot depends on the fine-tuning data, which should be collected in the robot's own embodiment and action space, and the model can fail on objects, scenes and instructions far from its training data. RoboTwin randomized success is lower because the fine-tuning data contain clean scenes only.
How to try it
- Open the model card for
vernerobotics/v0.5-baseand submit the access request. Read the license first. - Download the weights and minimal inference package once approved.
- For research comparisons, use the LIBERO or RoboTwin checkpoints in simulation.
- For your own robot, collect demonstrations in its action space and fine-tune, either with your own code or on Verne's platform.
- Run a risk assessment, with emergency stop and limits, before any physical test.
What this means for what you build
If you do robot learning research, V0.5 is a strong, documented baseline with per-task tables and episode counts, which is rarer than it should be. If you build products, the practical questions are license terms, the lack of fine-tuning code and how the model behaves on your robot, which no benchmark answers. Our coverage of related world-model and policy releases includes GreenWAM, LIBERO-PERM personalization results, the FineArt LeRobot dataset and Generalist's Gen-1.5.
Limits of this post
- We read the three model cards and Verne's website. We did not read the V0.5 paper or run the model.
- Comparison rows come from Verne's cards, not from our own leaderboard check.
- Access is gated, so we did not inspect the weights or license text beyond Verne's summary.
Related reading
- Ai2 ManiFlow-TB: an open robot action model
- GreenWAM open robot world model
- LIBERO-PERM personalized manipulation benchmark
- FineArt LeRobot dataset
- Generalist Gen-1.5 robot foundation model
Details are accurate as of November 5, 2026. Model cards, access terms and leaderboard rows may change. See Verne Robotics for current information.