RobotEra Releases VPP2 Weights for Open Robot Control
The open release lowers the cost of experimentation, but a 32.26% RoboDojo success rate keeps VPP2 out of production.
What was released
RobotEra released VPP2’s model weights on October 10, alongside code, configurations and evaluation guides. The world-action model predicts a visual trajectory for a task before connecting that plan to robot actions.
The release includes standalone video-prediction models and paired video-action checkpoints for RoboDojo and LIBERO. Teams can now run a local evaluation instead of waiting for a hosted commercial API.
Benchmarks are not uptime
VPP2 leads the official RoboDojo-Sim leaderboard with a score of 39.26 and a 32.26% success rate. Its paper also reports an 18.5 percentage-point advantage over the strongest baseline on real-world, zero-shot ALOHA manipulation tasks.
That is meaningful research progress, not a production guarantee. A leaderboard does not measure our objects, sensors, safety constraints, control-loop latency or the operational cost of a failed physical attempt.
What we would do
Virtual Arc would put VPP2 into a limited internal pilot now, but only in shadow or offline evaluation beside the existing system. We would measure per-task completion, latency, recovery behaviour and failures outside the expected distribution.
The model would not directly own equipment control. Deterministic motion limits, state validation and emergency stops would remain outside the model until results on real hardware justified the migration effort.
VPP2 is worth touching, but only as an evaluation branch, not as a shortcut to production. Downloadable weights, an MIT-licensed implementation and published evaluation recipes reduce vendor lock-in and let a robotics team measure the model on its own hardware. The leading RoboDojo-Sim result still comes with a 32.26% success rate, however, which is nowhere near the reliability we would accept for control of physical equipment. Virtual Arc would run VPP2 in an isolated test environment behind deterministic safety constraints, with every action recorded and replayable. We would not migrate an existing control architecture until it showed a material task-level gain on our workloads, acceptable control-loop latency and bounded failure behaviour on real robots. The open release buys a credible experiment; it does not buy permission to deploy.