Curr-0: Enabling Humanoid Robots to Perform Continuous Full-Body Dexterous Operations

Achieving fine manipulation while a humanoid robot is moving has long been an unsolved challenge in embodied intelligence, because the robot's stance, torso pose, foot contact, and hand forces are tightly coupled.

Traditional pipelines separate locomotion and manipulation into two independent modules: the robot first walks to a target, stops, and then performs the hand task. This works on constrained factory lines but fails in real-world settings where body posture directly influences reachability and force application.

Curr-0, released by Current Robotics, addresses this by training a single end-to-end policy—named Single Policy —that simultaneously controls locomotion, whole-body pose coordination, and precise hand actions on a 70+ degree‑of‑freedom humanoid platform.

The internal architecture consists of three layers. The top layer interprets task goals and language commands, the middle layer coordinates full-body motion and maintains stability, and the bottom layer handles hand‑object interaction. These layers operate within a single closed‑loop policy rather than a sequential pipeline, allowing real‑time adjustment of posture and hand movements during motion.

Training data comes from HumanEx , a self‑developed full-body exoskeleton system that captures 21,000 hours of authentic human behavior, including 2,800 hours of whole-body demonstrations. HumanEx records multi‑dimensional signals such as full-body pose, joint trajectories, hand motions, proprioception, EMG, and environmental interactions without requiring a robot body, enabling data collection in factories, labs, and offices.

Current Robotics treats this dataset as core infrastructure for embodied AI, arguing that data scale should be measured in "human task hours" rather than "robot deployment hours," allowing continuous growth as more human tasks are recorded.

To overcome the scalability bottleneck of physical testing, the team is building a multi‑physical‑modal interaction world model that simulates vision, proprioception, and force feedback. This world model supports strategy evaluation, post‑training, and pre‑deployment verification. The proposed Human‑in‑the‑World‑Model framework lets humans intervene directly in the simulated environment, instantly feeding corrected data back into training.

Overall, Curr-0 is presented not as an isolated model release but as a stage in Current Robotics' broader embodied‑intelligence stack that links data acquisition, model training, scalable evaluation, and iterative deployment.