Chinese Team Cited in π0 Unveils CurrentWorld-0 Interactive World Simulator
Current Robotics, whose work was cited in Physical Intelligence's π0, released CurrentWorld-0, an interactive world simulator that integrates cross-embodiment, multi-view, and force-tactile prediction for robot training.
The team's founder, Zhu Yichen, previously served as head of the Embodied Intelligence Department at Midea Group and has published over 40 top-conference papers with more than 5,200 citations. CurrentWorld-0 is the team's latest work following Curr-0, a model for whole-body fine manipulation of humanoid robots.
Robot world models often originate from video generation models, but large-scale pre-training on successful examples can cause the model to "correct" failed actions into successful outcomes, distorting evaluation. CurrentWorld-0 emphasizes action controllability: the world must change according to the robot's actions, and even if the robot makes a mistake, the model must not compensate for it.
To handle different robot embodiments, CurrentWorld-0 does not dictate a unified action format but preserves each embodiment's action subspace, allowing the model to predict subsequent object and scene changes from different action inputs. It also maintains synchronized multi-view records of the same event, preventing positional drift across camera views. For contact-heavy tasks like picking up chips or peeling cucumbers, the system incorporates force and tactile prediction to determine whether the gripper actually grasped the object or whether the knife truly applied pressure.
The simulator also supports human teleoperation intervention. When a policy enters a potential failure state, an operator can take over from the current state, or save and roll back the state to explore alternative actions. Each intervention generates new interaction trajectories that can be used for policy post-training.
Current Robotics states that while robot forms continue to diversify, all robots must ultimately be tested in the real world. CurrentWorld-0 is intended as the final checkpoint before that reality, where unfinished actions can be rehearsed, failed, and corrected.