In March 2026, Unitree open-sourced UnifoLM-VLA-0, a Vision-Language-Action model enabling the G1 humanoid to understand natural language task commands, interpret camera input, and execute physical actions without task-specific programming. Unitree released model weights, inference code, fine-tuning system, and 50,000 real trajectories under Apache 2.0 license. This is not a research paper. It is a deployable foundation model for robot control, publicly available to any G1 customer.
What a VLA Model Does
A standard industrial robot is programmed with explicit trajectories: move to X, close gripper, move to Y. The programmer specifies every step. A VLA model operates differently: the engineer says “pick up the red cup and place it in the bin on the left,” the model processes the visual scene, interprets the instruction, plans the manipulation, and executes. No explicit trajectory programming required. New task commissioning drops from 1-4 weeks to hours.
Architecture
UnifoLM-VLA-0 uses a transformer architecture with a vision encoder pretrained on internet-scale image-text data and a diffusion-based action decoder generating smooth, physically consistent motion trajectories. Input: text instruction, RGB image stream, current joint state. Output: joint position target sequence executed at 500 Hz by the G1’s low-level controller. Published benchmark results: 54 of 59 successful task completions (91.5%) on the VIRAL benchmark, above several proprietary VLA systems from US companies on the same evaluation suite.
Dex3-1 Hands: The Dexterity Layer
The benchmark results were achieved on a 29-DOF G1 with Unitree’s Dex3-1 dexterous hands, 7 DOF per hand, enabling pinch grip on objects as small as 5 mm. UnifoLM-VLA-0’s dexterous manipulation capabilities require the Dex3-1 configuration. With the standard parallel gripper, the action space is limited to two-finger grasping and performance on dexterous tasks drops significantly. Buyers evaluating the VLA model must include Dex3-1 hands in their procurement to realize the published capability.
Tokyo Haneda Airport Deployment
In April 2026, a G1-D wheeled variant was deployed at Tokyo Haneda Airport for baggage handling with Japan Airlines, the first commercial humanoid deployment in a Japanese international airport. Japan Airlines operates under JCAB safety regulations for automated equipment. Passing those requirements validates Walker S2 for regulated commercial service environments outside China, more rigorous than most demo deployments.
IPO and Pricing
Unitree filed for A-share IPO on the Shanghai Stock Exchange in March 2026, requiring audited financials, capacity disclosure, and governance standards of a listed company. G1 pricing: base $16,000, EDU $43,900+, G1 with Dex3-1 from $43,900. G1-D wheeled variant estimated $35,000-55,000. 16 configurations available on Unitree commercial portal. UnifoLM-VLA-0 commercial deployment license: negotiated per scale.
Competitive context: Top 40 China Robot Rankings 2025. AgiBot comparison: AgiBot A2 vs G2.
Sources
- Unitree: UnifoLM-VLA-0 open-source release (March 2026)
- Drones Plus Robotics: Unitree at CES 2026 new humanoid benchmarks
- ZMProbots: Unitree G1 complete guide 2026
- Botinfo.ai: Unitree G1 pricing all 16 configurations 2026
