Robotera’s open-source VPP2 world action model tops the RoboDojo benchmark and scores 58.5% zero-shot on a real ALOHA robot vs 40% for pi0.5
Partly confirmed
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 2 of 5
- Last verified
- 11 October 2026
Your AI and this story
- GPT-6 Astra163 days after its cutoff
- Claude Opus 5.5102 days after its cutoff
- Gemini 3.8 Flash193 days after its cutoff
- Grok 4.7132 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 102 days before it.
Key facts
- RoboDojo (42 dual-arm tasks; generalization, precision, long-horizon, memory, open-vocabulary): average success 32.26%, average score 39.26; first in generalization, precision and memory. Strongest baseline in its paper: GPT-6 Astra (post-training evaluation) 22.48% / 28.97 (QbitAI via Pandaily)
- Real robot: 58.5% average over 10 zero-shot task types (pick, place, stack, fold, pour) on ALOHA vs 40% for pi0.5; best on 9 of 10. LIBERO-Pro 45.0%, LIBERO-OOD 63.9%
- Architecture: event-level video pre-training on Wan2.1-I2V-14B; 8-second predictions distilled to a single step (~0.12 s); a 0.9B-parameter Action DiT learns actions while the video model stays frozen (LoRA only); ~0.22 s per action chunk
- Adding a VLM planner lifted selected long-horizon RoboDojo tasks from 27.6% to 57.6%
- Code open-sourced on GitHub. Competing claim: Simate said its Simate-beta topped RoboDojo in September
- Robotera runs robots in 10+ logistics centers with China Post and SF Express, but VPP2 itself is not reported as deployed at scale
What happened
VPP2 is a “video prediction policy”: it first imagines a video of the task being done, then reads actions off that prediction. Robotera reports state-of-the-art results on a new academic simulation benchmark and on real dual-arm hardware without extra training data. The figures are from the company’s paper as summarised by QbitAI and Pandaily; we did not read the paper or the GitHub repository directly.
Why it matters
It is a strong open-source data point for world-model-based robot control. It also shows a general frontier LLM (GPT-6 Astra) used as a baseline in robot-manipulation leaderboards, where specialised models still beat it.
Sources
1 source from 1 site. Numbers match the chips in the text.
1 source: 1 press
Press
Changes
- Filed (lab/press check)