Humanoid robots look best in a tightly cut demo. The real fight is uglier: collect real actions, train models across messy tasks, and make the same machine useful when nobody is filming.
Unitree’s 2026 releases show that shift. The Hangzhou company still ships attention-grabbing movement clips, but it is also dumping manipulation datasets, imitation-learning tools and more general robot models into the open. China’s humanoid boom is moving from choreography toward work that can be repeated.
Dancing is hard. Loading a dishwasher is harder.
Unitree first got famous with quadrupeds and athletic humanoid routines. Its 2026 company timeline lists an autonomous martial-arts show by 24 G1 robots in February, then roller skating, ice skating and voice-driven motion. Coordinated whole-body control is genuinely difficult. It still doesn’t prove a robot can clear a table, recover from a bad grip or finish a factory job without a handler.
Manipulation is a different sport: contact, object variation, occlusion, timing, and balancing while your arms do something useful. Unitree’s answer is data. On March 5 it launched UnifoLM-WBT, a whole-body teleoperation collection gathered on real robots in open environments, with rolling updates promised. Teleoperation means a human drives the robot to show a task; those recordings train policies that try it alone.
The Hugging Face collection makes this concrete. G1 datasets cover stuffing clothes into a washer, sorting fruit, making a bed, cleaning a living room, pushing a wheelchair and loading a dishwasher. Individual sets often run to hundreds of episodes, sometimes with different dexterous hands. That matters because these are full sequences, not isolated poses. A wider library is how you test whether a model generalises instead of memorising one kitchen.
One model, many hands
Unitree’s open-source page now lists UnifoLM-WLA-1.0, a general-purpose humanoid foundation model with 6 billion parameters. The company says it trained on about 2,500 hours of real-robot data and can coordinate 64 tasks across desktop and whole-body work, with one model spanning two-finger grippers and several five-finger hands.
Those are Unitree’s numbers, not an independent bake-off. The public write-up also doesn’t prove every task survives an unfamiliar room. The design bet is still clear: one model, many tasks, many end effectors — not a custom controller for every chore.
The tooling around it is half the story. The open-source portal lists manipulation datasets, an imitation-learning stack adapted from LeRobot, simulation tools, reinforcement-learning examples and robot SDKs. That is a path from recording to training to running on metal. Researchers get a faster lab. Unitree gets a platform story: sell the body, then become the place people build physical-AI skills.
Not only humanoids
Unitree hasn’t gone all-in on two legs. On February 24 it launched the As2 quadruped for industrial and professional work. Specs on the box: 90 N·m peak joint torque, more than four hours unloaded, IP54 splash protection, and over 13 km of travel with a 15 kg payload. Pandaily repeats those claims. Battery life always depends on terrain and load, but the product makes the point: inspection dogs, humanoids and arms can share perception and training progress even when the chassis differs.
What would actually impress us
Hundreds of demos are a start, not a labour force. Homes and factories throw lighting, floor surfaces, weird object shapes and bad collisions that no dataset fully covers. A robot that finishes slowly and safely beats one optimised for a viral clip.
The evidence that matters next is independent success rates, performance in new rooms, recovery after mistakes, and how often a human still has to intervene. Clear licenses and reproducible baselines matter too if outsiders are going to compare models honestly.
Unitree’s 2026 work is worth watching for the infrastructure pile-up: real-robot data, generalist models, open tools. Polished martial-arts videos are marketing. Whether these machines do everyday work is still the question that counts.