UnifoLM-VLA-0: The Architecture Behind Unitree's Robot

🤖🧧 The Chinese New Year robot which is doing Kung-Fu just open-sourced its brain.

Unitree Robotics, the startup behind the G1 humanoids that danced on CCTV, released UnifoLM-VLA-0: the Vision-Language-Action model that drives their humanoid.

And the architecture is fascinating for anyone working in ML or robotics.

This is NOT a classic VLM fine-tune on robot data. It's a 3-phase pipeline, each with a specific role:

Phase 1. Backbone Qwen2.5-VL-7B from Alibaba. A solid generalist VLM as the starting point.

Phase 2. Continued Pre-training → UnifoLM-VLM-Base The VLM is re-trained on robotics-specific tasks: 2D/3D detection, spatial reasoning, trajectory prediction, task decomposition.

This is where the model learns "physical common sense" with Forward & Inverse Dynamics constraints:

→ Forward: predict s(t+1) given s(t) and a(t). "What happens if I do this?" → Inverse: infer a(t) given s(t) and s(t+1). "What action leads to this state?"

Where s is the state and a is the action at time t. In VLA, this is what separates understanding from imitation.

Phase 3. Action Head (2 parallel branches) From the same VLM-Base checkpoint, two variants: • VLA-Base: 340h of real robot data on the G1, 12 manipulation tasks • VLA-LIBERO: trained on simulation, 98.7% on LIBERO benchmark (claimed by Unitree, surpassing π0 and GR00T)

What fundamentally changes compared to classic VLAs:

📦 Action Chunking. 25 actions predicted in 1 single inference, not step by step. 25× less network latency.

🌊 DiT Flow Matching. The action head is a Diffusion Transformer that denoises in 4 Euler steps. Random noise → coherent robot trajectory.

One single checkpoint. 12 tasks. Deployed on the G1 (1.32m tall).

Note: the Forward/Inverse Dynamics code is not in the open source repo. It's part of VLM pre-training (Phase 2), which wasn't published. LIBERO results and comparisons are Unitree's claims from their project page.