Galaxea has unveiled Fast-WAM, a new research breakthrough in world models for robotics.
As another milestone in World Action Models (WAMs), Fast-WAM moves beyond the traditional and inefficient “imagine first, act later” paradigm. By fundamentally rethinking the underlying model architecture, it tackles one of the longstanding bottlenecks in embodied AI: latency.
The breakthrough quickly sparked discussion across the AI community. Saining Xie, an assistant professor at New York University and a key contributor to diffusion transformers (DiT), also gave the work a nod, recommending Fast-WAM alongside LeWorldModel, the latest work by Turing Award laureate Yann LeCun.
The attention from leading figures in AI reflects how directly Fast-WAM addresses one of the most critical challenges facing the deployment of embodied world models.
For years, the prevailing assumption has been that WAMs must generate future video frames during inference, forcing robots into an inefficient “think first, act later” workflow. With Fast-WAM, Galaxea demonstrates that the power of a world model lies in its ability to model the world through video, rather than in generating video at inference time.
This insight removes a major computational burden, reducing single-step latency to just 190 milliseconds. Fast-WAM delivers more than 4× faster inference while maintaining state-of-the-art (SOTA) performance. Its arrival marks an important step toward real-time robotic control with world models, paving the way for the large-scale industrial deployment of embodied AI.
Paper: https://arxiv.org/abs/2603.16666
Project Website: https://yuantianyuan01.github.io/FastWAM/
1. Core Technical Breakthroughs in World Models: Training and Inference
As a pioneer in embodied AI, the Galaxea team approaches established technical directions—including vision-language-action (VLA) models and world models—with scientific rigor, seeking to uncover their underlying principles through carefully designed experiments.
Most mainstream WAMs follow an “imagine first, act later” paradigm. Given an observation, the model first runs a video diffusion module through iterative denoising to predict several future frames, then generates actions based on those imagined visual observations. Although intuitive, this approach has a major practical drawback: video generation is computationally expensive. Repeated denoising iterations can push inference latency into the hundreds of milliseconds, making it difficult to meet the real-time response requirements of robots.
Fast-WAM was designed to separate training from inference, two processes that are often tightly coupled.
- Training — joint modeling: Fast-WAM retains video co-training, allowing the visual module to learn physical dynamics and spatial interaction patterns.
- Inference — skipping generation: During task execution, Fast-WAM bypasses the redundant prediction branch. Rather than performing iterative denoising, it extracts latent world representations in a single forward pass, capturing the underlying physical dynamics.
This decoupling demonstrates a deeper understanding of how world models work and reflects the team’s commitment to maximizing algorithmic efficiency.
2. Experimental Results: 190 ms Inference Latency and More Than 4× Faster Inference
Performance ultimately comes down to experimental evidence. Fast-WAM delivers compelling results across the RoboTwin 2.0 bimanual manipulation benchmark, the LIBERO long-horizon task benchmark, and a real-world towel-folding task.
The team conducted three groups of controlled experiments, including variants that retained the video generation branch, generated video before actions, or removed video training. Together, these experiments demonstrate the advantages of the Fast-WAM approach.
- More than 4× faster inference: Fast-WAM achieves a single-step inference latency of just 190 milliseconds, compared with approximately 800 milliseconds for conventional WAMs. This substantial reduction in latency makes Fast-WAM better suited to the demands of real-time robotic control.
- SOTA performance among leading world models: On the RoboTwin 2.0 bimanual manipulation benchmark and the LIBERO long-horizon task benchmark, Fast-WAM achieves success rates on par with the strongest world models, even without generating future video during inference.
- Real-world validation: In the challenging towel-folding task, Fast-WAM demonstrates strong physical modeling capabilities when interacting with deformable objects, highlighting its practical potential for complex real-world manipulation.
With Fast-WAM, Galaxea has fundamentally rethought the world model approach. By removing the redundant future-video generation branch, it addresses a major limitation of world models in real-time control. More importantly, this shift has implications for the broader field of embodied AI beyond inference speed alone.
Fast-WAM challenges the long-held assumption that world models must generate future video. It demonstrates that physical understanding can be decoupled from high-latency generation. By delivering leading performance with an inference latency of just 190 milliseconds, Fast-WAM helps remove deployment barriers for world models in latency-sensitive applications such as robot arm manipulation and industrial automation. This opens up new possibilities for advanced policies that were previously constrained by computational cost and response time to move beyond the laboratory and toward large-scale industrial deployment.
Fast-WAM is more than another advance in algorithmic efficiency for Galaxea. It also reflects the team’s broader technical philosophy. Rather than blindly following prevailing paradigms or simply scaling up computational resources, Galaxea seeks to uncover the fundamental principles of embodied intelligence through rigorous analysis of underlying mechanisms and to find the most fundamental path forward for its evolution.
Explore the Fast-WAM technical report:
0 条评论