The race to build useful humanoid robots is shifting from better hardware to smarter AI.
Chinese robotics company Unitree has introduced UnifoLM-OminiA-0.3, an embodied AI model that combines speech, vision, navigation, reasoning, and physical manipulation into a single control system for humanoid robots. The company says the model is designed to help robots carry out household, wellness, and care-related tasks without switching between separate AI systems.
If that approach proves reliable outside carefully staged demonstrations, it could bring humanoid robots closer to practical use in homes, hospitals, and elder-care settings, where tasks and human instructions constantly change.
In a video demonstration released by Unitree, a humanoid robot completes a variety of domestic and care-related tasks. It was able to place a cushion on a sofa, identify colors, count medicine boxes, retrieve a requested medication package from a shelf, sort clothes into a laundry basket, load dishes into a dishwasher, and adjust a hospital-style bed.
The demonstration also highlights the system’s ability to react while a task is already underway. As the robot adjusts the bed, a user tells it to stop, and the robot immediately halts the action instead of waiting until the task is complete.
Building on Unitree’s broader AI strategy
UnifoLM-OminiA-0.3 is the latest addition to Unitree’s UnifoLM embodied AI program.
The company previously introduced UnifoLM-WMA-0 in 2025 as an open-source world-model-action framework to help robots predict how environments change after physical interactions. That was followed by UnifoLM-VLA-0 in early 2026, a vision-language-action model for general-purpose humanoid manipulation released alongside training code and model weights.
Unitree has also expanded its collection of whole-body teleoperation datasets featuring humanoid robots performing household activities such as cleaning and object handling. Many of those tasks appear again in the new OminiA-0.3 demonstrations.
While many of the individual tasks shown have already been demonstrated by other humanoid robot companies, Unitree’s main claim is that one AI model can manage the entire process—from understanding spoken language and recognizing objects to planning actions, navigating spaces, and controlling the robot’s body.
That reflects an emerging trend across the robotics industry, where developers are moving away from building separate task-specific programs and toward general-purpose AI systems that can adapt to changing environments and human interaction.
What it could mean
If Unitree’s approach proves reliable outside controlled demonstrations, it could make humanoid robots more practical for homes, hospitals, and elder-care facilities, where instructions often change, and environments are far less predictable than factory floors.
The challenge, however, is that demonstrations do not guarantee consistent real-world performance. Homes are filled with clutter, pets, changing layouts, and unexpected interruptions that remain difficult for robots to handle. Success will depend not only on smarter AI models but also on whether they can safely perform these tasks day after day without constant human oversight.
Whether UnifoLM-OminiA-0.3 represents a genuine leap forward will depend on performance outside carefully controlled demonstrations. As robotics companies race toward general-purpose humanoids, the real test is no longer whether robots can complete isolated tasks—but whether they can safely adapt to the unpredictable environments people live in every day.
More News: China’s humanoid robot race is shifting from flashy demos to real-world deployment. Meet AGIBOT’s new A3 Ultra, an eight-hour humanoid designed for factories and service work—and why its focus on uptime could matter more than acrobatics.
