NEWS / SEP.2026
Delta Intelligence introduces Δ₀, a model for controlling humanoid robots
Delta Intelligence introduced Δ₀, its intelligence and control model for humanoid robots, on 28 September 2026. Its founder, Xiaojian Ma, is targeting energy and infrastructure in the near term, while the results published by the company concern internal tests in domestic settings.
Delta Intelligence targets hazardous occupations with its Δ₀ model
Delta Intelligence unveiled Δ₀ on 28 September 2026, a model designed to coordinate humanoid robots’ movements and object manipulation. The launch reported by Leiphone includes domestic demonstrations and evaluations conducted by the company.
In written responses published the same day by Humanoids Daily, founder Xiaojian Ma identifies energy and infrastructure as near-term priorities. The domestic market is being considered for the longer term.
Ma cites underground facilities, offshore sites and environments subject to extreme heat or humidity. He believes their operators would be more willing to pay for machines capable of keeping technicians away from difficult or dangerous work.
Coordinating hands and points of support
In a demonstration published by Delta Intelligence, the robot opens a dishwasher. Its left hand braces against a cabinet while its right hand pulls the door. The robot adjusts its posture and points of support to follow the door’s rotation while maintaining its balance. This scene illustrates the connection between manipulation and balance.
Delta Intelligence describes an architecture in which visual observations, the instruction and the robot’s state feed into a world-action model. This model anticipates how the scene will evolve and produces motion commands. A learned controller converts these commands into targets for the joints of the hands, arms, torso and legs.
Delta reports 69 degrees of freedom, each corresponding to an independent movement coordinated by the system. This number describes the configuration demonstrated. Humanoids Daily specifies that the demonstrations use, among other platforms, a modified Unitree G1, controlled by the Δ₀ intelligence and control layer.
To pretrain the main model, called the “brain,” Delta reports using more than 10 000 hours of data combining first-person human videos and whole-body movements. The controller also learns from movement data.
For long sequences, Delta says that intermediate goals come from a higher-level model, an agent or a person. Δ₀ handles the physical transitions between these steps, such as repositioning or maintaining a grip. Fully autonomous mission planning from a single instruction remains to be demonstrated.
Xiaojian Ma also describes a system under development, with a head equipped with cameras and a back-mounted module housing computing hardware and a battery. This system aims to bring Delta’s models to different humanoid platforms.
A mirrored kitchen tests Δ₀’s adaptability
In a separate internal test in a mirrored kitchen layout, Delta evaluates Δ₀ on the dishwasher task. The sink and dishwasher switch sides relative to the training scene. For this task, Delta reports 4 successes out of 20 attempts, or 20 %, with the system trained on the original demonstrations.
Delta then collects attempts accompanied by targeted human corrections and performs reinforcement learning post-training. This learning adjusts behaviors based on the outcomes of actions. After this additional training, the company reports 13 successes out of 20 attempts on the same dishwasher task, or 65 %, under the same stated evaluation conditions.
The twenty attempts in each series constitute a small sample. The reported improvement concerns adaptation to this kitchen after additional training. Δ₀’s reliability at the industrial sites targeted by Xiaojian Ma remains to be established.