DeTa Intelligence Unveils Delta-0, a Native Foundation Model for Humanoid Robots
On September 28, DeTa Intelligence unveiled Delta-0, a native foundation model for humanoid robots. The model uses a "brain-cerebellum-force-position" architecture and converts inputs such as vision, language and body state into unified whole-body action targets.
In a one-take demo, a humanoid robot powered by Delta-0 planned 32 steps over 210 seconds to complete nine household tasks across bedroom, living room and kitchen, including tidying a bed, putting items away, loading a washing machine and dishwasher, discarding spoiled fruit and sitting on a sofa. Another robot equipped with a high-degree-of-freedom dexterous hand picked up a vinyl record, started a record player and moved to the music. DeTa said the demo shows long-horizon autonomy, whole-body coordination and agility comparable to Figure Helix 2.5.
According to the company, the "brain" directs limb movements using vision-dominated multimodal sensing, the "cerebellum" handles whole-body coordination and converts commands into high-frequency motor control, and force-position hybrid control balances position and contact force. Delta-0 reuses existing upper-body manipulation and cross-embodiment data along with human whole-body action data, and uses a "real-sim-real" learning and evaluation pipeline with real interaction feedback, human correction and reinforcement learning for post-training and failure recovery.
