RobotEra has put VPP2 first on RoboDojo-Sim’s non-agent table, a result that gives robot builders a concrete new policy to inspect but not a reason to assume reliable physical task completion. VPP2’s 39.26 average Score and 32.26% success rate lead that table. That is a meaningful separation in a setting designed to measure robot-task performance. It also makes the remaining question unusually clear: can the same system preserve that advantage when simulation gives way to a physical evaluation?
On its October 8 RoboDojo simulation leaderboard, the benchmark lists VPP2 as a RobotEra contribution and places it first among non-agent entries. The next listed non-agent system recorded a 35.34 Score and a 29.64% success rate. Readers who first encounter the result through QbitAI should start with that narrow comparison: it is a lead on RoboDojo-Sim’s non-agent table, not a global robot ranking or a real-world result.
That distinction matters because the benchmark’s two figures answer different questions. A lead in partial progress can show that a system moves more effectively through pieces of a task, while a binary success figure asks whether it reaches the required finish. RobotEra’s entry leads on both measures, but neither number turns a one-third success rate into evidence that every task is completed reliably. The value of VPP2’s result lies in the signal it offers for the next experiment, not in a claim that the experiment is already over.
RoboDojo-Sim gives VPP2 a useful but limited lead
RoboDojo combines a partial-progress Score with binary task success across five dimensions. A policy may make meaningful progress without completing a task, and a completed task counts differently from a run that stops short. The 39.26 Score therefore shows VPP2’s average standing in the benchmark’s progress measure; the 32.26% figure is the share recorded as successful under its binary measure. Reading those values together is more demanding than treating either as a headline on its own.
The gap to the next non-agent entry is useful for a practical reason. It shows VPP2 did not merely edge another listing on a single measure. Its advantage appears in both the average Score and the success-rate column, with the next entry at 35.34 and 29.64%, respectively. For teams deciding what to examine, that creates a defensible reason to look at VPP2’s approach. It does not establish that the margin will carry into a different benchmark, a different machine, or a physical setting.
A separate EastFrontier look at VA-Bench’s robot control loop in simulation offers useful context for readers weighing what simulation measures can and cannot settle. The immediate lesson for VPP2 is not that its lead should be discounted. It is that the result should be read for what the table actually records: simulated progress and task outcomes within RoboDojo’s stated setup.
That framing protects against two opposite errors. Dismissing a simulation lead would overlook a visible performance difference among the non-agent entries shown on RoboDojo-Sim. Elevating it into a blanket statement about robot capability would erase the condition that produced the result. The row tells readers that RobotEra’s submitted system performed best among those entries; it does not answer whether it can deliver the same outcome outside that table.
VPP2’s video-to-action design creates the next question
The technical case for examining VPP2 is more specific than the ranking. The VPP2 preprint, first submitted October 7, describes a system built around staged video pretraining, planning, and action generation. This places the leaderboard result beside a stated design path: learn from video in stages, use planning, then produce actions. The approach is relevant because RoboDojo’s table measures what the submitted system did, while the preprint describes how its authors say the system is organized.
The authors also report 58.5% average success across ten zero-shot ALOHA categories. That result is worth keeping separate from the RoboDojo row. It is a preprint research result reported by the authors, not independent deployment evidence, and it covers a different stated evaluation. Folding the figures into one larger claim would make both harder to understand. One is a current RoboDojo-Sim leaderboard position; the other is an author-reported average over named ALOHA categories.
The public repository makes VPP2 easier to inspect: it is available under an MIT license. Yet the repository retains planned-release items for the pretrained backbone and large-scale pretraining components. This creates a practical trade-off. The visible code can help readers and technical teams understand the policy’s direction, while the unavailable planned components limit what can be treated as a fully reproducible version of the reported work.
For readers tracking the commercial stakes around embodied AI, China’s embodied-AI financing tally is useful background on why the connection between a policy and a body deserves scrutiny. VPP2’s public materials make the distinction concrete: a simulation result can be compared now, while the complete materials needed to reproduce its larger pretraining path are not all released.
A RealWorld Evaluation Would Put VPP2 Under a Harder Test
The strongest reason to keep the VPP2 result in proportion is simple. The RoboDojo RealWorld leaderboard that was checked does not list VPP2. That absence does not negate the simulation lead. It identifies the unresolved boundary between a promising policy result and a physical demonstration that can be examined on the same benchmark family.
A physical result would also put the released checkpoint through a sharper comparison. If VPP2 were evaluated across three embodiments, readers could see whether the same checkpoint survives the change in embodiment. Reproducible artifacts would let others establish what version was tested. Failure data would make it possible to distinguish a system that falls short in a recurring way from one whose misses are scattered across tasks.
This is where the 32.26% success rate becomes most useful. It is not a verdict on VPP2; it is a baseline that makes a future physical result legible. A higher Score in simulation can be encouraging, but a three-embodiment physical evaluation would reveal whether the system’s partial progress translates into completed work under the harder condition. A lower or uneven physical result would be just as informative, because it would expose where the simulated advantage stops traveling.
RobotEra’s next test is therefore precise rather than abstract: whether the same released checkpoint can survive a three-embodiment RoboDojo physical evaluation with reproducible artifacts and failure data. Until then, the simulation lead is a credible reason to watch VPP2 closely, and an incomplete answer to the question that matters most.