X-Humanoid releases open-source Robo-ValueRL framework for humanoid precision manipulation
X-Humanoid released Robo-ValueRL as an open-source offline-to-online reinforcement-learning framework for humanoid precision manipulation. The framework uses value estimation to guide VLA policy improvement and targets millimeter-level manipulation tasks such as chip insertion.
Evidence notes
- Robo-ValueRL links value estimation, quality-conditioned policy learning and online residual adaptation for humanoid manipulation workflows.
- The project page reports 240 hours of offline data, more than 3,000 online rollouts and 86% final success on millimeter-level chip insertion.
- The release adds open-source code, model and dataset access for robot-learning developers working on precision-critical humanoid tasks.
Company context
Beijing-based humanoid robotics innovation center developing the Tiangong humanoid platforms, embodied-intelligence systems and open developer infrastructure for humanoid manipulation, locomotion and industrial pilots.