Latent-action robot models report new policy and world-model results
October 26: RynnWorld-Latent adds a cross-embodiment world model pretrained on about 553.5 million frames from 23 datasets, with synthetic mid-training gains.
A 4B latent-action robot policy and a follow-on world model now report strong simulation and real-robot results.
RynnVLA-Latent is a 4B robot policy that predicts continuous, 3D-grounded latent actions instead of native controls. RynnLAM produced 745.52 million descriptors from 30.1 thousand hours of human, robot, and simulation video. The policy reached 98.8% on LIBERO and 84.7% on LIBERO-Plus without benchmark-specific adaptation, with 51.9% average success on five VLABench tracks and 88.6% on RoboTwin 2.0, and the highest observed success among compared methods on six real-world tasks. On October 26, RynnWorld-Latent was reported as a cross-embodiment video world model conditioned on those latent actions, pretrained on about 553.5 million frames from 23 datasets. Digital teleoperation, latent-action synthesis, and synthetic mid-training improved evaluated success over real-demonstration-only training.
- RynnLAM turned 30.1K video hours into 745.52 million latent-action descriptors.
- The 4B policy scored 98.8% on LIBERO and 84.7% on LIBERO-Plus without adaptation.
- Average success was 51.9% on five VLABench tracks and 88.6% on RoboTwin 2.0.
- It led compared methods on six real-world tasks.
- RynnWorld-Latent pretrained on about 553.5 million frames from 23 datasets.
- Synthetic mid-training raised policy success on LIBERO and LIBERO-Plus.
Timeline
dated events in the reporting · oldest first- Oct 24, 2026RynnVLA-Latent 4B policy and RynnLAM scale reported · Hugging Face daily papers
- Oct 26, 2026RynnWorld-Latent world model and pretraining scale reported · Hugging Face daily papers
What each outlet adds
beyond the shared facts- Hugging Face daily papersOctober 24 policy scores, descriptor count, and real-task ranking.
- Hugging Face daily papersOctober 26 world-model frame count, 23 datasets, and synthesis gains.
- Products
- RynnLAMLIBEROLIBERO-Plus
Coverage
oldest first · one row per article · UTC- · 6d ago
Hugging Face daily papersFirst report
RynnVLA-Latent pretrains a 4B robot policy on 3D latent actions
A 4B policy pretrained on 745.52M descriptors from 30.1K hours of video posts strong LIBERO, VLABench, and RoboTwin scores.
- · 4d ago
RynnWorld-Latent conditions a world model on universal latent actions
Pretraining spans 23 datasets and about 553.5 million frames, then adapts to target robots without a shared action space.