Galaxy General's Wang He Outlines WAM Roadmap Toward Embodied AI's 'ChatGPT Moment' in 2028
At the WRC 2026 (World Robot Conference) main forum in Beijing, Wang He, founder and CTO of Galaxy General (Galbot) and researcher at Peking University, delivered a concrete answer to a question hanging over the entire industry: how should embodied AI's "ChatGPT moment" be defined, and when will it arrive? His answer: a three-year roadmap to 2028, built on the World-Action Model (WAM) paradigm.
Defining the 'ChatGPT Moment'
Wang He's definition is unusually specific: a robot achieves zero-shot generalization of 70-80% success rate on common skills it was never specifically trained on, and ordinary users without algorithm backgrounds can teach it new behaviors via simple post-training to reach near-100% success in their specific scenario.
> "With today's data accumulation and continued breakthroughs in post-training, Galaxy General will reach embodied AI's ChatGPT moment in 2028. Once that level is reached, industry deployment scale and speed will rise sharply — by the end of the 15th Five-Year Plan, tens of thousands of humanoid robots across industries."
He judged the current state of embodied foundation models as equivalent to the "ChatGPT-2 era," with the leap to "GPT-3.5 to GPT-4" capability targeted around 2028.
AstraBrain WAM 0.5: Fusing VLA and World Models
Wang He framed the paradigm debate:
- VLA (Vision-Language-Action): the mainstream route since Google RT-2; strength is absorbing action-labeled data — but only action-labeled data.
- World Model: can absorb massive unlabeled video, but can only imagine futures, not directly command every joint.
- Wang Xingxing (Unitree): the tipping point is a robot completing 80% of tasks in any unfamiliar environment from voice commands — "two to three years at best, five to ten at worst."
- Huang Qingqiu (Moqi Intelligence): data is the industry's foundational constraint — a chicken-and-egg problem.
- Wang He: "2028, embodied AI's ChatGPT moment."
- Fully autonomous tennis playing by a humanoid (March, a world first), which Elon Musk called "embodied AI's AlphaGo moment" — backed by over 10,000 years of simulated training.
- Spring Gala walnut spinning — dexterous-hand reinforcement learning on objects whose tightness constantly changes.
- World-first high-dynamics dexterous pen spinning — frequent pen-finger contact that simulators struggle to render faithfully.
Galaxy General's answer is WAM (World-Action Model), which trains on both action-labeled robot data and internet-scale unlabeled human video under one unified objective. Key milestones:
1. WAM was first proposed by Galaxy General in a March 2025 paper accepted to ICCV 2025; by August 2026, 171 follow-up papers on world-action models had appeared. 2. AstraBrain WAM 0.5 models the future in latent space — stripping lighting and task-irrelevant textures — and already performs cross-scene, cross-embodiment, multi-task work. 3. NVIDIA's head of embodied AI publicly called WAM "a lifelong game for robotics" in a Sequoia interview.
AstraBrain-WBC 0.5: The General 'Cerebellum'
If WAM is the brain, WBC (Whole-Body Control) is the cerebellum. WBC 0.5 was trained on 2 billion frames of human motion-capture data, with a demonstrated scaling law: loss keeps decreasing as data grows past 100,000 hours. Demonstrated tasks include teleoperated pet feeding and litter cleaning, high-dynamics dance (handstands, human-robot co-dancing) from the Galbot ET1 launch, and emergency/industrial/retail support. Compared to NVIDIA's general cerebellum, Wang He claimed lower latency and better precision and stability.
The Robot Trio: One Brain, Three Bodies
| Robot | Type | Key scenarios | | --- | --- | --- | | Galbot ET1 | Bipedal humanoid | Real-time multimodal interaction, high-dynamics dance | | Galbot G1 | Wheeled semi-humanoid | Breakfast making, garment folding, replanning under disturbance | | Galbot S1 | Heavy-load wheeled | CATL production line, 7×24h industrial tasks |
At WRC competitions, Galbot won gold in all three flagship categories — home, supermarket, and restaurant — with 100% gold-medal coverage: restaurant completed autonomously in 6 minutes 55 seconds with zero errors; supermarket finished order picking, random restocking, and misplaced-item return within 20 minutes.
The Data Pyramid: 'Galaxy Star Tree'
| Layer | Data type | Scale | | --- | --- | --- | | L1 Internet data | Video, images, text | General crawling | | L2 Human data | Hand manipulation | 1M hours (500K bare-hand + 500K UMI head-mounted) | | L3 Synthetic data | Simulation + rendering | In-house simulator | | L4 Teleoperation | Training fields in 6 provinces/cities | Industry-wide collection | | L5 Real-robot return | Autonomous deployed work | 80K hours (healthcare, CATL, pharmacies) |
The goal: scale data volume 10x by 2028. At WRC, Galaxy General also released the Astra Set head-mounted dual-hand UMI capture device, with over a thousand units reportedly deployed across industries.
Commercial Deployment
> "Since December 2024, our robots have picked over 1 million medicine boxes across nearly 100 pharmacies in Beijing, Shanghai, Guangzhou, Shenzhen and other cities. Mean time between mechanical failures has reached one and a half years, and we aim for three to five."
On August 9, Beijing Haidian unveiled China's first provincial/ministerial-level key laboratory in embodied AI foundation models, co-built by Galaxy General, Peking University, and CATL, with Wang He as director and an academic committee led by CAE academician Gao Wen.
WAM-TTT: Post-Training from Plain Videos
WAM-TTT (Test-Time Training) pushes post-training门槛 to a minimum: customers just record ~10 minutes of work video with a head-mounted camera — no labels, no teleoperation, no action data — and the base model is adapted into a scene-specific model. "Every industry can put the model to use," Wang He said.
Industry Crosscut: Diverging Timelines at WRC 2026
Technical Showpieces
Observations and Open Questions
Wang He did three concrete things: pushed WAM from paper to production line, turned the "ChatGPT moment" slogan into a quantifiable metric (70-80% zero-shot + 100% post-trained per scene), and hard-bound commercialization to 2028 with real data infrastructure behind it.
Open questions:
1. Is 2028 too optimistic? Going from GPT-2 to GPT-3.5 took about four years (2019-2022); embodied AI aims to compress that to two. 2. Is WAM truly the next paradigm? 171 follow-up papers and NVIDIA-level endorsement are strong signals, but boundaries versus VLA, world models, and behavior cloning need independent validation. 3. Can video-only post-training scale reliably? WAM-TTT's low barrier raises engineering questions about customer data quality and avoiding corrupting the base model.
Worth tracking over the next 6-12 months: WAM adoption by 3-5 leading embodied-AI companies, the arrival of AstraBrain WAM 1.0, and the real-world data behind the three WRC gold medals.