DreamZero

Headlines · 3

1. Oct 19
   EVO-WAM lets world action models self-train on verified imagined trajectories

   Researchers behind EVO-WAM report a self-training loop in which a world action model generates complete imagined video-action trajectories, a vision-language model checks whether the task was completed, and an inverse dynamics model checks that the generated actions match the imagery before verified segments are used for fine-tuning. The team reports that Cosmos3's average success rate rose from 26.9% to 68.0% on seven unseen RoboTwin 2.0 simulation tasks and from 20.0% to 76.7% on three unseen long-horizon real-robot tasks with 10 trials each, while DreamZero improved from 28.5% to 46.4%. An ablation found 43.7% fourth-round success with vision-only filtering versus 68.0% with the action-consistency check, and the authors note gains were not monotonic, with a dip in round three.

   36氪 人工智能 · 🔥 0
2. Oct 14
   Robots learning from human video: Robot Era opens VPP2 weights, Dyna-2 scale test

   On October 10, Robot Era released two sets of video model weights for VPP2. According to an article by Embodied Observer published via TMTPost, VPP2 splits robot skill learning into three stages — predicting how a task unfolds, compressing it into a visual plan, and learning robot actions — and only the first two were opened. The piece reviews several routes for putting human video into robot training: Dyna-2 compared pretraining on 1,000 to 1,000,000 hours of first-person video, and on a key-turning task only the 1,000,000-hour group succeeded, 9 times out of 10 tries; Nvidia's DreamZero raised average task progress from 38.3% to 54.3% after adding 12 minutes of action-label-free human video; τ0-WM used about 27,300 hours of heterogeneous data. The article also notes a gap between predicted video and executable action: DreamZero's oven task failed after the predicted plan skipped opening the door, and in RoboWM-Bench simulation, Wan 2.6 videos converted into actions succeeded 50% of the time at closing a drawer, 20% at picking an object and 0% at placing it in the drawer.

   钛媒体 · 🔥 0
3. Oct 8
   Stanford's Fei-Fei Li team open-sources OpenWAM world-action model framework

   A Stanford team including Fei-Fei Li, Jiajun Wu and Ehsan Adeli published OpenWAM, an open framework for composable world-action models that aims to disentangle the many ways video prediction and action generation can be coupled in this emerging class of robot policies. OpenWAM starts from Alibaba's open Wan2.2-5B video model, continues pretraining on roughly 3.34 million robot-and-human interaction videos (nominally about 14,600 hours) for 14 days on 32 B200 GPUs, then joins a 5B video expert with a 2B action expert via a shared Mixture-of-Transformers architecture with four configurable interaction orders (VTA, ATV, Joint, Decoupled). The authors report 98.6% average success for VTA across the four LIBERO subsets and 92.1% / 91.9% average success for VTA and Joint on three real dual-arm Franka FR3 tasks, plus a LIBERO-Long-CF counterfactual dataset; training, evaluation and composition code and pretrained video weights are open-sourced.

   36氪 人工智能 · 🔥 0

Experience and discussion from the community

Share my DreamZero experienceAsk about DreamZero