AnswersPublished23 sources
Robotera’s VPP2 Leads RoboDojo-Sim, but Large-Scale Deployment Remains Unproven
Robotera reported VPP2 at 32.26% success and 39.26 average score on RoboDojo Sim, plus 58.5% zero shot success on real ALOHA tasks; these are benchmark results, not proof of large scale deployment. VPP2 learns to predict video before generating robot actions.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Robotera reported VPP2 at 32.26% success and 39.26 average score on RoboDojo Sim, plus 58.5% zero shot success on real ALOHA tasks; these are benchmark results, not proof of large scale deployment.
- VPP2 learns to predict video before generating robot actions. Its reported RoboDojo lead is specific to the simulation leaderboard, separate from RoboDojo’s real world evaluation.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What does the October 10, 2026 report reveal about Tsinghua University-backed Robotera’s open-source VPP2 world action model, including its. Article summary: The October 10 report presents VPP2 (Video Prediction Policy 2) as an open-source world action model with a reported RoboDojo-Sim lead and stronger real-robot zero-shot results than pi0.5—not as proof of large-scale comm. Topic tags: general, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, click
Robotera’s VPP2 pairs video prediction with robot action generation. On RoboDojo’s simulation leaderboard, the company reported a 32.26% average success rate and a 39.26 average score, with category leads in generalization, precision and memory. A separate real-robot test reported 58.5% zero-shot success across ALOHA tasks. The findings are notable, but they do not show that VPP2 is deployed at scale. 2
6
What the RoboDojo results measure
RoboDojo, led by the University of Hong Kong’s MMLab with academic collaborators, evaluates robot manipulation across 42 dual-arm tasks. The tasks test generalization, precision, long-horizon execution, memory and open-vocabulary instruction following. The benchmark reports both task success and an average score, so its results capture more than a single completion rate. 5
6
Robotera’s reported results put VPP2 first on the simulation leaderboard at the time of the announcement: 32.26% average success, 39.26 average score, and first place in generalization, precision and memory. This is a leaderboard result, not a claim that VPP2 leads every RoboDojo evaluation: the benchmark’s simulation and real-world leaderboards are separate. 2
6
In a reported post-training comparison, GPT-6-Astra scored 22.48% success and 28.97 average score. That puts VPP2 ahead by 9.78 percentage points in success rate and 10.29 points in average score in the cited comparison. These figures should be read in the context of the reported evaluation, not as a universal ranking across all robot tasks. 6
There is also a timing caveat around “first place.” RoboDojo’s leaderboard history lists an earlier Simate-beta result, followed by VPP2’s October 8 entry. Rankings describe a particular leaderboard and date; they are not timeless claims about all versions, settings or real-world performance. 12
What “without additional data” means
Robotera said it achieved the benchmark result without additional data or agent-based reinforcement self-improvement. That claim should not be read to mean VPP2 was trained without a substantial video-training process: descriptions of the method say it builds on a pretrained Wan2.1-I2V-14B model and continues pretraining with diverse robot-manipulation, human-activity and general videos. 25
9
11
The distinction matters: the no-extra-data claim concerns the reported benchmark result and add-on self-improvement, while the described method still uses a curated video dataset to develop the model’s prediction ability. 9
11
How VPP2 trains prediction before action
VPP2’s central design choice is to develop video prediction first, make it faster, and then add action generation. The stated aim is to use predicted future scenes to inform movement while reducing the risk that action training will erode the video model’s learned generalization. 9
11
- Learn to predict manipulation. Starting from Wan2.1-I2V-14B, the first stage continues video pretraining with robot-manipulation, human-activity and general footage. The videos are described in detail and organized around meaningful events, so the model learns to predict actions across different tasks.
9
11
- Compress prediction into a planner. The next stage trains for a fixed eight-second prediction horizon and uses consistency distillation to compress multistep generation into a single-step visual planner. Reported video-prediction latency is about 0.12 seconds.
18
- Generate actions from predicted futures. The final stage adds a 0.9-billion-parameter Action DiT, connected to the video model through a mixture-of-transformers design. The described training setup initially freezes the video backbone and uses LoRA adaptation for the action component, aiming to preserve the video model’s predictive representations.
4
The reported action-generation time is about 0.1 seconds for a two-second action chunk, alongside roughly 0.12 seconds for video prediction. These are reported experimental timings, not a hardware-independent speed guarantee. 11
21
Real-robot and LIBERO results
On a real ALOHA dual-arm robot, VPP2 reportedly achieved 58.5% average zero-shot success across 10 task types, compared with 40% for Physical Intelligence’s pi0.5—an 18.5-percentage-point difference in that evaluation. These physical-robot results are distinct from the RoboDojo simulation leaderboard and should not be treated as directly interchangeable with it. 2
6
The project’s repository reports 45.00% on LIBERO-Pro and 63.93% on a modified LIBERO-OOD protocol. The modified-protocol qualification is important when comparing the OOD figure with results from other setups. 17
Open-source release is not proof of deployment
VPP2’s official GitHub repository documents the implementation and release process. Its release notes say video-model weights became available on October 10, while RoboDojo history-conditioned video-training code was announced on October 12—evidence of a staged release rather than a single complete release date. 3
22
The available evidence supports reported benchmark performance, real-robot testing and open-source availability. It does not establish that VPP2 itself is operating at scale in commercial logistics deployments. Company deployments, even if independently verified, would not by themselves show that those systems run this particular model.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Robotera’s VPP2 Leads RoboDojo-Sim, but Large-Scale Deployment Remains Unproven"?
Robotera reported VPP2 at 32.26% success and 39.26 average score on RoboDojo Sim, plus 58.5% zero shot success on real ALOHA tasks; these are benchmark results, not proof of large scale deployment.
What are the key points to validate first?
Robotera reported VPP2 at 32.26% success and 39.26 average score on RoboDojo Sim, plus 58.5% zero shot success on real ALOHA tasks; these are benchmark results, not proof of large scale deployment. VPP2 learns to predict video before generating robot actions. Its reported RoboDojo lead is specific to the simulation leaderboard, separate from RoboDojo’s real world evaluation.