Unitree Robotics has introduced UnifoLM-OminiA-0.3, a new embodied-AI model designed to allow humanoid robots to carry out a range of household, wellness and care-related tasks using a single control system. The Chinese robotics company describes the system as a model for real-time, omni-modal interaction and whole-body mobile manipulation. It combines information from sources including spoken instructions and visual observations, allowing a robot to understand a request, move through its surroundings and manipulate objects without switching between separate task-specific models.
In an accompanying demonstration, a Unitree humanoid is shown completing several activities in domestic and care environments. These include picking up a cushion and placing it on a sofa, identifying colours, counting medicine boxes, retrieving a requested box from a shelf and sorting clothing into a laundry basket.
The robot is also shown loading a plate into a dishwasher and operating an adjustable bed. During the bed demonstration, the user verbally instructs the robot to stop while the task is in progress. The robot responds to the interruption and ends the action, illustrating Unitree’s focus on interaction that continues while the robot is working rather than only before a task begins.
One model for several stages of a task
The individual activities shown in the video have previously appeared in demonstrations by other humanoid robot developers. The main point of Unitree’s release is therefore not any single household task, but the company’s claim that one model can coordinate several stages of physical work.
These stages include understanding spoken and visual information, recognising objects, deciding what action is required, positioning the robot and controlling its arms and hands. Unitree says the model is intended to execute tasks autonomously and maintain stable performance when interruptions or other disturbances occur.
This approach reflects a wider development in humanoid robotics. Manufacturers are increasingly attempting to connect language and visual models directly to robot actions, rather than programming a separate sequence for every object, room or activity.
For applications in homes and care environments, this ability would be necessary because robots cannot rely on the highly structured conditions found in many factories. Objects may be moved, instructions may change and people may interrupt a robot while it is operating.
Part of the wider UnifoLM programme
OminiA-0.3 forms part of Unitree’s broader UnifoLM embodied-intelligence programme. In September 2025, the company released UnifoLM-WMA-0, an open-source world-model–action architecture intended to help robots understand physical interactions and predict how an environment may change following an action.
Unitree followed this in January 2026 with UnifoLM-VLA-0, a vision-language-action model for general-purpose humanoid manipulation. The company published training and inference code alongside model weights. The model was developed to combine visual and language understanding with the generation of physical robot actions.
Unitree has also released a growing collection of whole-body teleoperation data recorded with humanoid robots. The dataset includes activities in domestic environments and is intended to support the development of models that combine locomotion with object manipulation.
Several activities represented in these earlier datasets, including cleaning living spaces and handling household objects, also appear in the OminiA-0.3 demonstration.
