Robot Model SOTA Now Ships With a Nationality Qualifier

Figures attributed to companies, projects and papers are their published claims. The synthesis and judgements below are this author’s analysis. I read the release posts, the GitHub README, the leaderboard protocol and the paper abstracts; I have not run any of these models or robots.

Verne Robotics published V0.5 on 1 November and called it “an open-weight world action model built to accelerate the deployment of general-purpose robots in industrial settings”. The claim attached to the benchmark table was narrower: on LIBERO and RoboTwin 2.0, V0.5 is “the leading open-weight model developed by a US company” at the time of publication.

Four qualifiers stack up before the word “leading”: open weight, US company, two named benchmarks, as of a date. None of them is dishonest, and the evaluation behind them is more transparent than most. But a claim that needs four qualifiers is telling you about the state of the field, not about the model.

The same week, Meituan’s robotics team submitted Meituan-Robotics-0 to arXiv, built on 147 million vision-language samples and 15,060 hours of robot trajectories across 12 embodiments. Two open robot foundation models in seven days, from two of the largest economies, both positioned as the base layer everyone else fine-tunes. The model layer has become a commodity, and the announcements now read like commodity marketing: pick a benchmark, pick a sub-league, take the top slot.

The table where the leader is third

Verne’s GitHub README publishes a RoboTwin 2.0 comparison sourced from the leaderboard of 30 October 2026. It reads, in order of the clean scene column:

Model	Clean	Randomized
PICA-WAM	90.7%	74.7%
GO-3 Nano Preview	88.5%	81.5%
EWAM	82.2%	72.1%
ViGAR	82.0%	67.0%
Verne V0.5	81.0%	73.4%
UniWAM	75.1%	67.6%
π0.5	70.7%	46.0%
π0	46.4%	16.3%

V0.5 is third by the board’s own ranking, the mean of clean and randomized success, behind GO-3 Nano Preview and PICA-WAM, and fourth on clean scenes alone. Publishing a table that ranks you third is more honest than most release posts manage. It is also the reason the qualifier exists. On the benchmark that varies the scene, Verne leads nothing except a geographic subset.

The LIBERO numbers need their own table, because this is the benchmark that gets quoted in headlines:

Model	Average of 4 suites	Long suite
ABot-M0.5	99.4%	98.4%
Verne V0.5	99.3%	99.2%
VPP2	98.8%	98.0%
Cosmos Policy (NVIDIA)	98.5%	97.6%
X-VLA	98.1%	97.6%
Fast-WAM	97.6%	95.2%
π0.5	96.85%	92.4%
π0	94.15%	85.2%

Six models inside 1.8 points. Verne reports a 95% interval for its own 99.3% average of 98.8% to 99.7%, and ABot-M0.5’s 99.4% sits inside it. By Verne’s own statistics, first and second place on LIBERO average are not separated. The Long suite is different: 99.2% against 98.4% and 92.4% is a real gap, and the only ordering in the table with an argument behind it, since Long chains sub-goals and the failures are compositional rather than perceptual.

So: competitive but not clearly ahead on a saturated benchmark, better than the previous open generation on multi-step tasks, mid-pack on the only test that randomizes the world. For the architecture side, the earlier piece on world action model conditioning covers what these networks condition on.

Randomization is the column that discriminates

RoboTwin 2.0’s randomized split changes clutter, textures, lighting and table heights. Its leaderboard protocol fixes training at 50 clean demonstrations per task across 50 tasks, evaluates 100 trials per task on an Aloha-AgileX, and ranks by the mean of clean and randomized success. The shared harness makes the table comparable.

The clean to randomized drop is the number I would put on a datasheet. PICA-WAM loses 16.0 points (90.7 to 74.7), π0.5 loses 24.7, π0 loses 30.1, EWAM loses 10.1, GO-3 loses 7.0, and V0.5 loses 7.6. The models that hold up best under randomization are not the ones with the highest clean scores. PICA-WAM wins the clean column by 2.2 points over GO-3 and then finishes behind it once the lighting changes. If your deployment involves a real table, real lighting and other people’s clutter, the randomized column is the ranking you want, and it produces a different order than the press release.

Meituan-Robotics-0 complicates the picture further. The paper reports 93.6% success on RoboTwin 2.0, above every figure in Verne’s table, and fifth place in the RoboDojo VLA track. Those are two different boards, and MR0 carries a much larger pretraining corpus than the fixed 2,500 demonstration budget assumes, which is what the leaderboard’s separate co-train track exists to acknowledge. I cannot close the gap between 93.6% and 90.7% from the two posts alone, and neither release settles it. Treat MR0’s figure as the strongest claim of the week and the one most in need of an independent rerun.

The gap between 99% and 24%

The number that should worry anyone planning a deployment is on neither leaderboard. VICTR, posted to arXiv the same week, post-trains π0.5 for in-context action generation and tests it on unseen, out-of-domain real-world tasks. The fine-tuned π0.5 baseline succeeds in 6.2% of attempts. With progress-informed retrieval of demonstration chunks, success rises to 23.8% and task progress from 28.9% to 55.5%.

The model family that posts 96.85% on LIBERO posts 6.2% when the task falls outside its training distribution and the scene is real. That is the honest scale of generalization in late 2026: leaderboard success in the high nineties, novel-task success in the low twenties, and more than double the progress rate from a retrieval mechanism that never touches the weights.

CS-Pi, accepted at CoRL 2026, attacks a related gap. A flow-matching policy can satisfy constraints that were never in its demonstrations, such as avoiding an obstacle it has not seen, but only if enforcement happens at sampling time. Gradient guidance fails in a specific way: it steers the noisy state using the policy’s running estimate of the clean action, and when the policy is uncertain between passing left or right of the obstacle, that estimate is the average of the two, which goes straight through it. The paper’s draw-project-renoise sampler reaches 81% safe success on a shelf pick-and-place against 55% for plain projection, and 55% on a Franka against 40%.

None of this is a criticism of the models. It is an argument about which number predicts whether a robot works on Monday morning. A 99% simulation score is evidence that a policy learned the benchmark’s distribution. A 24% out-of-domain score is evidence about the deployment.

The model does not include the controller

The technical section I keep returning to in Verne’s post is not about the transformer. It is about servo tracking.

Verne’s robots mostly use quasi-direct-drive actuators rather than harmonic drives with joint torque sensors. QDD is backdrivable, handles contact well and costs much less at fleet scale, but gravity, payload changes and inter-joint coupling show up as motion nobody commanded, so accuracy has to come from the command rather than the gains. The stock YAM controller sends each target as a position with zero target velocity, so damping pulls against every motion: the arm trails by 52 to 60 ms at every speed, and median end-effector error grows from 9 mm to 26 mm. Verne’s controller plays targets from a queue filled ahead of time and feeds forward joint velocity, acceleration and torque, which brings lag to 4 to 12 ms and keeps median error between 2.2 and 3.5 mm. Same policy, 20 runs per controller, both arms. At 0.5 m/s, 52 ms of lag is roughly 26 mm of travel, the same order as the tracking error and as the tolerance on a connector insertion.

Put that next to the tables. A policy at 99.3% and one at 98.8% differ by half a point. A controller that removes 45 ms of lag changes where the gripper is by more than a centimetre at production speed. The published success rate is a joint property of the model and the execution stack underneath it, and almost no paper separates the two. The first question about a four point gain from a new architecture should be whether the baseline ran on a stock position controller. The piece on end effector poses and force works the same seam: what the policy emits decides what the controller can recover.

ActPrune attacks the same seam from the other side. Visual token pruning usually happens after the vision encoder has processed the whole observation, too late to matter for closed-loop control. ActPrune decides earlier, using the previous action chunk as a geometric prior: the third-person view keeps patches along the projected end-effector trajectory, and the wrist view prunes patches that stay unchanged once action-induced camera motion is compensated. Training-free, up to 1.61 times faster with 59.5% fewer vision encoder FLOPs, negligible success change on LIBERO, VLABench and four real tasks.

That is the practical shape of the problem. V0.5 emits 32 targets at 30 Hz, about 1.07 seconds of motion, and the controller runs at 500 Hz underneath. Everything between those two rates is engineering, and it decides deployment outcomes.

What this means if you are choosing a model this quarter

Do not pick on LIBERO. The top six are inside 1.8 points and the intervals overlap.

Ask for clean and randomized numbers separately, and ask what changed between them. A single success rate hides the measurement that predicts whether a policy survives a different table height.

Price the adaptation, not the pretraining. Verne claims a new task, or a new step of a longer task, can be learned from under two hours of demonstrations and corrections. That is a vendor claim, and it assumes you can collect good demonstrations.

Read the license and the hardware bill. V0.5 is 8.2 billion parameters in bf16, 17 GB of weights, and the README asks for an NVIDIA GPU with at least 24 GB of VRAM, 32 GB of RAM and 35 GB of disk, plus a Hugging Face login after accepting the model license. Verne’s license is its own agreement, not an OSI approved one. Open weight here means downloadable after a form, the norm across this generation and a different thing from open source; the roundup on cable robots and open robot weights tracks how fast that tier is filling up.

Measure your own loop before blaming the model. If the arm trails the command by 50 ms, no policy will look as good on your robot as it did in the paper.

What to watch

Whether MR0’s 93.6% reproduces on the official RoboTwin board under the co-train label, and whether the gap turns out to be a training budget difference or an evaluation difference. A number a third party cannot reproduce is a press release, not a result.

Whether the open weight tier keeps its national qualifiers. This week’s pattern is a US release defining leadership inside a US subset, and a Chinese release that does not bother with the qualifier because it does not need one. That holds only as long as the US models stay within a couple of points.

Whether anyone publishes the failure rate on a task the model was not trained for. That number is currently buried in a paper about retrieval, and it is worth more to a deployment team than the whole LIBERO table.

Sources

• Verne Robotics, Introducing V0.5: An Open-Weight World Action Model, 1 November 2026
• Verne V0.5 GitHub repository, README with benchmark tables, checkpoints and hardware requirements
• Verne V0.5 model licence
• RoboTwin 2.0 leaderboard, benchmark protocol and submission policy
• Meituan Robotics Team, Meituan-Robotics-0: A Vision-Language-Action Foundation Model for Desktop Manipulation, arXiv:2611.08577, 5 November 2026
• Macesanu et al., VICTR: Value-Informed In-Context Learning for Vision-Language-Action Models, arXiv:2611.08698, 5 November 2026
• Moon et al., ActPrune: Action-Grounded Visual Token Pruning for Efficient Vision-Language-Action Models, arXiv:2611.08827, 6 November 2026
• Bhattacharya, Kroemer and Tulsiani, CS-Pi: Constraint-Satisfying Flow Policy Sampling, arXiv:2611.08654, CoRL 2026