Unitree has announced a partial open-source release of UnifoLM-WLA-1.0, an embodied intelligence model built on approximately 2,500 hours of real-robot data. One operator, working nonstop, would need 104 consecutive days to accumulate that amount. The model is roughly 6B parameters, with a single set of weights covering 64 tasks — 54 tabletop manipulations and 10 whole-body or mobile manipulation tasks. That data scale is rare in embodied AI, where most public models rely on simulation or web video with real-robot hours often in the low hundreds.
On September 26, 2026, Unitree founder Wang Xingxing pushed the narrative further in a talk titled "From Mechanics to Intelligence — An Evolutionary Theory of the Embodied Future" at the 5th Global Digital Trade Expo. The same week, Unitree released *part* of UnifoLM's weights, code, and datasets — and "part" is the key word.
Key points
- Model components: UnifoLM-ER-1-4B (spatial reasoning, Qwen3-VL-4B base, 5M+ embodied samples covering point prediction, detection, multi-image reasoning, 2D trajectories, 3D detection, and spatial QA); ER-Flow (scene-change and action representation on top of ER-1); and the WLA model itself, which adds an MMDiT action expert decoding continuous robot actions.
- Dynamic region prediction: optical flow extracts future scene changes, compressed via VQ-VAE into fixed-length mask tokens so the VLM predicts which regions will move after interaction begins — a mental rehearsal before acting.
- Discrete action learning: residual vector quantization splits a unified action space into end-effector pose, hand joint, and lower-body joint tokens aligned to shared timesteps, enabling whole-body coordination tasks like towel folding and bed making in one model.
- Benchmarks (vendor-reported): Unitree claims ER-1-4B leads participating open models on 7 of 16 multimodal spatial benchmarks, and scores above proprietary GPT-6 Astra on some spatial tests (Where2Place, EmbSpatial). No independent replication or released evaluation scripts exist yet.
- Released: UnifoLM-ER-1-4B weights, ER-Flow weights, UniBot-V1 Challenge Dataset
- Pending: UnifoLM-WLA-Base weights, post-training code, Unitree-WBT-Dataset, Unitree-Manipulation-Dataset
- Unitree GD01 rideable mech: claimed world's first mass-produced rideable mech, ~3m tall, ~500kg with a rider, priced above 3 million RMB; positioned as an "off-road vehicle of robots" for outdoor terrain.
- AgiBot (Zhiyuan Robotics): its 20,000th general-purpose embodied robot rolled off the line and was delivered to Chimelong, with 300+ units doing tours, shows, and educational programs; the 10k-to-20k jump took about six months.
- Qiyuan Robotics: launched personal robots Q1 and T1 on September 20, both priced at 19,999 RMB.
- Whether the WLA-Base weights and post-training code are actually released — the only hard test of Unitree's "open source" claim
- Third-party replication on non-Unitree hardware (the published setup includes two-finger grippers and various five-finger dexterous hands)
- Whether the community adopts Wang's 80%×80% yardstick — no public third-party evaluation currently uses it
- Whether real-robot data scale continues to grow past the currently scarce 2,500-hour level
What is released vs. what is not
This is the crux. The project page lists Code, Models, and Datasets as "Coming soon." A September 16 release checklist shows:
The messaging operates on two levels: the publicity layer claims 6B parameters, 64 tasks, and 2,500 hours of real data; the reproducibility layer delivers only the 4B perception/reasoning components and a challenge dataset. The full weights carrying those 64 tasks are not yet available to outside researchers. This gap is not necessarily false advertising — a hardware company open-sourcing all its manipulation capabilities would be commercially irrational — but it means the spatial reasoning portion is verifiable while the whole-body control portion, the hardest and most questioned part of embodied AI, is not.
Wang Xingxing's threshold: 80% × 80%
Wang offered a quantifiable definition of embodied AI's "ChatGPT moment": when robots can complete 80% of tasks via voice-driven embodied capability in 80% of unfamiliar scenarios, the industry hits its breakout inflection point — which he expects within a few years. He identified the biggest bottleneck as insufficient precision matching between AI model input/output and the real physical world: LLMs operate in digital space with minimal error, while robots face physical losses. He also judged that embodied AI and humanoid robotics are only three to four years old and remain in an early stage, years away from mass household adoption.
Same-month industry deliveries
Together, these suggest September 2026's theme is embodied AI shifting "from impressive videos to deliverable products," with three independent validation tracks: openness of the brain, mass deployment in tourism, and consumer availability.
What to watch
Sources
1. Unitree UnifoLM-WLA-1.0 official project page and release checklist check (September 16, 2026 status) 2. Studio Global, "Unitree UnifoLM-WLA-1.0: What's Released and What's Still Missing," 2026-09 3. NetEase, "人形机器人开卖不稀奇，具身大脑肯开源才靠谱," 2026-09 4. Cailianshe / STAR Market Daily, coverage of Wang Xingxing's talk at the 5th Global Digital Trade Expo, 2026-09-26 5. China Business Journal, "造车之外的新战场 车企竞速具身智能赛道," 2026-09-26 6. Yang Tianruo, "Tensor AI" talk at NetEase Future Conference, 2026-09-23