Unitree Robotics shifts to AI models, open-sources embodied foundation model UnifoLM-WLA-1.0 for home tasks

September 10, 2025 — Unitree has open-sourced a new embodied foundation model capable of driving robots to complete over 60 whole-body and desktop manipulation tasks, including organizing dishes, taking out the trash, and loading clothes into a washing machine.

According to the company, the model outperforms both domestic and international open-source models across multiple embodied reasoning benchmarks, and is comparable to leading closed-source models.

Since first releasing its proprietary embodied intelligence model in September 2025, Unitree has launched several models under the UnifoLM series. Following its IPO, the company plans to allocate nearly half of the raised funds to the development of intelligent robot models.

The Rise of the Robot "Butler"

On social media, users often question why humanoid robots are needed when factories already have precision robotic arms.

Wang Xingxing, founder of Unitree, addressed this at the 2026 World Robot Conference, noting that the biggest bottleneck in embodied intelligence today is the lack of generalization — the ability for robots to apply learned skills to new situations.

"In fixed scenarios, with sufficient data collection, many robot models can achieve 100% success. But if the object or environment changes even slightly, success rates drop significantly."

Generalization performance is therefore a key metric for evaluating embodied intelligence models.

UnifoLM-WLA-1.0: Key Specifications

• Parameters: 6 billion
• Training data: ~2,500 hours of high-quality real-world robot data
• Capabilities: 10 whole-body tasks + 54 desktop manipulation tasks
• Hardware compatibility: Two-finger grippers and multiple five-finger dexterous hands
• Key advantage: Cross-task and cross-end-effector generalization

Unitree has released real-world demonstration videos of all 64 tasks on its official website. Most tasks are household-related, including:

• Organizing shoe racks
• Cleaning the kitchen
• Sorting utensils
• Doing laundry
• Making the bed
• Retrieving food from a microwave

These tasks range from simple actions — like picking up a book and placing it on a shelf — to complex, multi-step operations, such as:

The robot walks to a coffee table, uses a five-finger dexterous hand to grab a water bottle, bends down, and throws it into a trash bag. It then picks up the bag with its right hand, transfers it to its left hand, walks to a pedal-operated trash can, steps on the pedal with its right foot, holds the lid open with its right hand, and deposits the bag with its left hand.

According to a July report by KPMG, while robots can already perform short, well-defined actions like grasping, placing, and opening doors, real-world scenarios often involve long-horizon, multi-step tasks with strong contextual dependencies — such as tidying a desk, cleaning a room, warehouse sorting, or factory assembly. These tasks require the model to continuously track object positions, occlusion states, task progress, and the downstream effects of prior actions.

How the Model Works

The UnifoLM-WLA-1.0 model consists of two core modules:

1. UnifoLM-ER-Flow (Multimodal Backbone): Responsible for understanding the world and determining what actions to take.
2. MMDiT Action Expert Module: Translates these understandings into executable, smooth motion commands for the robot.

The multimodal backbone integrates three tasks:

• Embodied reasoning
• Future dynamic region prediction
• Discrete action learning

This allows the robot to not only perceive its environment but also predict how its interactions will change it, and encode physical actions into a format the model can learn from.

The goal is to align visual input, language commands, predicted changes, and encoded actions within a shared semantic space.

Benchmark Performance

Unitree reports that its UnifoLM-ER-1 model — responsible for embodied reasoning — achieved 7 state-of-the-art results across 16 multimodal perception and understanding benchmarks, outperforming open-source models and rivaling closed-source foreign models. In benchmarks such as Where2Place, BLINK, and EmbSpatial, it scored higher than OpenAI's latest GPT-6-Astra flagship model.

Wang Xingxing stated that Unitree's overall R&D capabilities in embodied AI have placed the company among the global leaders. The company plans to further increase investment in embodied AI models, data collection, and real-world training.

"Robot intelligence is a top priority for us. I personally have devoted the most energy to training robot AI models. Globally, even major AI companies worry about falling behind the rapid pace of AI evolution. We maintain that vigilance as well."

He added: "My team and I must iterate, improve, and learn quickly. We need to propose innovative ideas and run experiments. Run enough experiments, and some will succeed — that's how we stay ahead."

From Hardware-First to Hardware-Software Synergy

For a long time, Unitree was perceived primarily as a hardware company, with a cautious approach to embodied AI investment.

In an August 2024 interview, Wang acknowledged that the technical roadmap for robot AI models was still unclear — unlike large language models. "You can't invest heavily when the technical path is fuzzy. If you burn through enormous resources without clarity, you're just running in place."

However, he also emphasized that hardware serves AI. "If you don't understand robot AI models, it's hard to build a good humanoid robot."

Since September 2025, Unitree has shifted its strategic focus from "hardware + cerebellum" to "brain + cerebellum + hardware," releasing multiple models and pursuing two parallel technical architectures:

• WMA (World-Model-Action)
• VLA (Vision-Language-Action)

Model Release Timeline (2026)

| Date | Model | Key Highlights | |------|-------|----------------| | January 2026 | UnifoLM-VLA-0 | Open-source VLA model; evolves from general image-text understanding to physically aware "embodied brain" | | February 2026 | UnifoLM-X1-0 | Industrial-grade model deployed in Unitree's own factory for tasks like motor assembly | | May 2026 | WVLA2.0 | Combines world model with VLA; supports long-horizon planning, multi-step prediction, and precise hand-eye coordination | | July 2026 | UnifoLM-OminiA-0.3 | Single model managing diverse tasks in home and elderly care scenarios; full multimodal interaction | | September 7, 2026 | UnifoLM-X2-1.0 | First real-time world-model-driven autonomous combat; breakthrough in instantaneous planning and dynamic interaction |

Source: Southern Metropolis Daily