Unitree Open-Sources UnifoLM-WLA-1.0: 6B Humanoid Robot Foundation Model With 7 Open-Source SOTAs
智东西 · wechat · 2026-09-18
Unitree has unveiled UnifoLM-WLA-1.0, a 6B-parameter general-purpose humanoid robot foundation model trained on 2,500 hours of real-robot data, and is open-sourcing the model, code, and datasets.
Capabilities
- Its reasoning base UnifoLM-ER-1-4B (built on Qwen3-VL-4B, 5M+ embodied reasoning samples) leads open-source models on 7 of 16 benchmarks, beating GPT-6Astra on Where2Place (82.0 vs 69.0) and EmbSpatial (88.9 vs 83.3).
- A single model handles 64 tasks — 54 tabletop and 10 whole-body operations — from folding clothes to taking out trash, supporting two-finger grippers and multiple dexterous hands.
Technical approach
- Interaction-centric world modeling: optical flow identifies regions an action will change, compressed via VQ-VAE into discrete tokens, so the model predicts "what will change" before acting.
- Action space is split into three streams (end-effector pose, end-effector joints, lower-body joints), discretized with RVQ into UnifoLM-ER-Flow, then decoded into continuous actions by an MMDiT Action Expert.
Moat: while the model is open, the real barrier is the 2,500 hours of real-robot data (equivalent to one robot collecting for 100+ days straight) and the ability to keep producing it — Unitree is aiming for an "Android moment" for embodied intelligence.
More from Embodied
- Baidu's Apollo Go wraps two-week public robotaxi test rides in Tung Chung, Hong Kong — Baidu_Inc · 2026-11-04
- imec neural probe integrates photonics directly into 130nm CMOS backend — jwt0625 · 2026-11-04
- Chris Paxton: interactive robot sim is the future but glaringly data-limited — Scobleizer · 2026-11-04
- DreamX: video diffusion models as universal latent action learners for robots — kaist-ai · 2026-11-04
- RoboTryOn turns human videos into robot demos, lifting success rate from 2% to 60% — kaist-ai · 2026-11-04
- INMO Future glasses: dual 1080p displays, 125g, $799 early-bird on Kickstarter Dec 10 — Scobleizer · 2026-11-04