Delta Intelligence has released the Delta-0 model, which is a native humanoid foundation model designed for full-body manipulation. This release occurred on September 28, 2026. The model is officially positioned as a brain-cerebellum-force/pose stack that integrates vision, language, and proprioception into unified target commands for whole-body actions. This enables bipedal creatures to move, maintain balance, and perform manipulations within a single continuous cycle, rather than relying on separate modules for locomotion and manipulation.
During a demonstration covered by LeiPhone, the Delta-0-based humanoid was able to take thirty-two steps and complete nine household tasks in approximately 210 seconds. These tasks were performed in bedroom, living room, and kitchen settings, including making the bed and gaming console, picking up a stuffed toy from the floor, loading laundry, placing a cup in the dishwasher, finding a spoiled fruit in the refrigerator, throwing it into a pedal bin, and sitting on a sofa.
Furthermore, the second arm with high degrees of freedom played a vinyl record and danced to music, serving as a test of precise loco-manipulation linking grasping, posture, and rhythmic whole-body movement.
Architectural Features and Training
The model's architecture clearly separates functions: the 'brain' translates multimodal sensory data into limb movement commands; the 'cerebellum' handles high-frequency instincts and balancing; and the hybrid force/pose control regulates contact, maintaining stance stability. The training process utilizes upper limb and cross-embodiment data, along with human whole-body motion capture augmented by real-world interaction feedback, human corrections, and reinforcement learning for failure recovery.
Delta Intelligence also describes a 'real-simulation-real' evaluation loop, which allows scenes to be recreated in simulation for navigation stress testing and retries before returning to hardware for skill transfer verification.
Developers' Goals
Founder and CEO Dr. Ma Xiaojiang defined the goal of developing robots capable of operating in human-scale spaces and solving long-horizon tasks, rather than merely imitating the human silhouette. Co-founder and Chief Scientist Huang Xiuyuan noted that large language models assist in planning steps but do not inherently provide contact stabilization, balance, or continuous force. The Delta-0 model is positioned as transferable to platforms such as Unitree and AgiBot, as partners continue work on scene adaptation and verification.