Seonghoon, a PhD student at KAIST and research intern at NVIDIA, explains DreamZero, a world action model for zero-shot robot policies, in a Robo Papers conversation with host Michael.
Seonghoon asks whether robots can generalize to new tasks the way people do, without collecting repetitive demonstrations for every task. He describes video models as a source of visual planning and proposes adding inverse dynamics to extract actions from predicted future states.
The team collected about 500 hours of teleoperation across 20 to 30 mock environments, encouraging varied, long episodes rather than repeated tasks. DreamZero is a 14-billion-parameter autoregressive model that jointly predicts video and actions, using language, visual observations, and proprioception.
The model runs in a closed loop: it predicts actions, the robot executes them, and fresh observations update its context. System and model optimizations cut action generation from five to six seconds to 0.15 seconds, or about seven hertz. In reported evaluations, it outperformed the tested VLA baselines on seen and unseen tasks, with 90% overall success across shirt folding, fruit packing, and table bussing after post-training.
Seonghoon argues that data distribution and diversity matter alongside quantity: collecting 50,000 useful robot hours would be difficult, while video prediction can help models learn from varied examples. In small transfer experiments, 10 to 20 minutes of unlabeled human or other-robot video improved unseen-task performance, and 30 minutes of diverse data helped adapt to a new embodiment.
Seonghoon says failures often come from inaccurate video prediction around object contact, while dexterous work may benefit from tactile and force signals. DreamZero also has limited video context for long-horizon tasks. The team is exploring smaller distilled models; the released model and training code are open source, though inference currently needs expensive GB200 GPUs.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Nadun and Juny, second-year PhD students at Georgia Tech, explain SAIL, their co-authored approach to faster imitation-learning policy execution, with RoboPapers host Michael.
Jason Ma, a robotics researcher at the University of Pennsylvania who worked on the paper at Google DeepMind, joins RoboPapers hosts Michael and Chris to discuss using vision-language models to assess robot task progress and data quality.
