Juno connects robot predictions to motion and action learning
In plain words
Juno is a framework that helps robot policies use predictions about what may happen next when choosing actions. It trains a compact scene summary to reflect motion around objects, then connects predictions to the part of the policy that generates actions.
In four simulated WidowX tasks, Juno had higher average task success than Qwen3GR00T, though carrot placement was nearly unchanged. The paper also reports results on a second simulation and on one physical robot setup.
The research
Robot prediction models can forecast what a scene will look like, but their features may not reflect movements a particular robot can control. How can those predictions help a robot act and remain useful when its surroundings change?
Juno trains one action-conditioned predictive model to represent robot motion, guide action generation, and adapt to new conditions.
Juno uses an action-conditioned predictive model across robot representation training, policy learning, and deployment adaptation.
It trains a compact scene state to track motion, connects predicted future states to action generation, and can update its world model from deployment interactions before re-aligning the policy.
In SimplerEnv’s four simulated WidowX manipulation tasks, Juno’s mean success rate was 68.5% versus 60.9% for Qwen3GR00T; episodes were scored successful or not, and carrot placement was nearly flat. In RoboCasa-GR1’s 24 simulated GR1 tasks, the unweighted mean was 59.6% versus 47.8%.
These results cover the reported simulation tasks and one physical robot setup; they do not establish performance across other robots, tasks, or datasets. Test-time adaptation results also use extra deployment rollouts and optimization, so they are separate from the frozen-policy evaluations.
Analysis
Deep
- Authors
- Yuchen Zhu, Chenyi Xu, Yulin Zhang, Gang Xu, Wentao Zhu
- License
- arXiv non-exclusive