Full specifications
6 of 14 fields stated| Field | Value | Source and caveats |
|---|---|---|
| Release | ||
| Released | 17 March 2026 | [A] "Submitted on 17 Mar 2026" D |
| Input and output | ||
| Model type | world action model | Project page, section heading: "A WAM That Drops Future Imagination at Test Time". B |
| Inputs | camera language | Paper: "We consider embodied policy learning from visual observations and language instructions." The cameras: "Images from multiple cameras are concatenated into a single image before being fed into the VAE." A |
| Controls | not stated | Note The paper names the output as action chunks, not a robot body part: "we introduce an action expert DiT for action chunk generation". A |
| Degrees of freedom controlled | not stated | |
| Size and speed | ||
| Parameters | 6B | Paper: "resulting in a 1B action expert and a total model size of 6B parameters". A |
| Control rate | not stated | Note The page gives an inference time, not a control rate: "Runtime: 190 ms on a single RTX 5090D V2 32GB GPU". B |
| Where it runs and what it is built on | ||
| On robot or cloud | not stated | Note Latency was measured on a GPU: "All latency numbers are measured on a single NVIDIA RTX 5090D V2 32GB GPU." The paper does not say where the model runs on a robot. A |
| Base model | Wan2.2-5B | Paper: "Fast-WAM is built on top of the video Diffusion Transformer (DiT) from Wan2.2-5B". A |
| Openness | ||
| Access | Open weights (download) | Hugging Face model repository, not gated: "This repository includes the original FastWAM checkpoints". C |
| Licence terms | not stated | Note The code repository carries the MIT licence ("MIT License"); the checkpoint page names no licence for the weights. C |
| Licence | not stated | Note The code repository carries the MIT licence ("MIT License"); the checkpoint page names no licence for the weights. C |
| The developer’s own claims | ||
| Learns new tasks | not stated | |
| Works across robots | not stated | Note The paper's real-world test used one robot: "Real-world towel-folding task on the Galaxea R1 Lite platform." A Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count. |
How to read the figures
- 50HzA
- The letter in the grey square after a figure points to the source the figure came from. Pages with several sources carry several letters.
Notes
- The Fast-WAM authors are affiliated with Tsinghua University and Galaxea AI.
- The released code carries the MIT licence.
Versions and notes in the key figures (7)
Model type
Project page, section heading: "A WAM That Drops Future Imagination at Test Time".
Controls
Note The paper names the output as action chunks, not a robot body part: "we introduce an action expert DiT for action chunk generation".
Inputs
Paper: "We consider embodied policy learning from visual observations and language instructions." The cameras: "Images from multiple cameras are concatenated into a single image before being fed into the VAE."
Access
Hugging Face model repository, not gated: "This repository includes the original FastWAM checkpoints".
Parameters
Paper: "resulting in a 1B action expert and a total model size of 6B parameters".
Base model
Paper: "Fast-WAM is built on top of the video Diffusion Transformer (DiT) from Wan2.2-5B".
Licence
Note The code repository carries the MIT licence ("MIT License"); the checkpoint page names no licence for the weights.
Sources
-
A
arxiv.
org/ · Retrievedhtml/ 2603. 16666v1 -
B
Primary source. yuantianyuan01.
github. · Retrievedio/ FastWAM/ -
C
huggingface.
co/ · Retrievedyuanty/ fastwam -
D
arxiv.
org/ · Retrievedabs/ 2603. 16666
One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.