Fast-WAM
World Action Model on a Wan2.2-TI2V-5B video backbone plus a 1B action expert, trained with a joint flow-matching objective over action chunks and future video latents. Keeps video modeling during training but predicts actions directly at inference, without generating future frames.
6B paramsopen weightsverified 2026-10-24paper →
Architecture
- Family
- world-model
- Parameters
- 6,000,000,000 (Authors state a total model size of 6B parameters (Wan2.2-5B backbone + 1B action expert), rounded.)src
- Backbone
- Wan2.2-TI2V-5B (video DiT, text encoder, video VAE) + 1B action expertsrc
- Action head
- flow-matchingsrc
- Action chunking
- yes
- Cameras
- min 1src
- Language input
- required
Compute requirements
No sourced requirements yet.
Training requirements: see also the LeRobot compute hardware guide, the authoritative source for LeRobot training envelopes.
Measured inference 0 runs · inference cost only, not task success
No measured runs yet. Submit one with its runtime, precision and camera count.
Robots running Fast-WAM 2 of 24 with evidence
| Robot | State | Tier | Verified | Detail |
|---|---|---|---|---|
| ROBOTIS OMX | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| SO-101 | partial | T1 · upstream documented | 2026-10-24 1d ago | Upstream docs only give a generic real-robot rollout command for so101_follower; no SO-101 training recipe, real-robot checkpoint or result is published. The only SO-101 checkpoint found is trained on simulated SO-101 data. |
No verified evidence, but both are available in LeRobot (ecosystem context): AgileX PiPER, ARX5, Bimanual SO-100/SO-101, EarthRover Mini Plus, HopeJR, I2RT YAM, Koch v1.1, LeKiwi, OpenArm, Reachy 2, reBot Arm B601-DM, SO-100, Trossen WidowX AI, Unitree G1, Universal Robots UR5e, UFACTORY xArm, XLeRobot.
Framework support
Checkpoints
- lerobot/fastwam_base · LeRobot-org base checkpoint (Hub license tag apache-2.0).
- lerobot/fastwam_libero_uncond_2cam224 · LeRobot-org LIBERO checkpoint, 2 cameras at 224 px (no Hub license tag).
- lerobot/fastwam_robotwin_uncond_3cam_384 · LeRobot-org RoboTwin checkpoint, 3 cameras (no Hub license tag).
- ZibinDong/fastwam_libero_uncond_2cam224 · Checkpoint used for the LIBERO results in the LeRobot docs (96.4% average over 2000 episodes).
- yuanty/fastwam · Released upstream checkpoints.
Related terms
Sources
- T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot docs describe FastWAM as a World Action Model that predicts actions directly at inference.
“FastWAM keeps video modeling during training, but uses direct action prediction at inference time instead of iteratively generating future observations.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/fastwam.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Data requirements: one or more cameras, optional state (proprio_dim), action, and a language task or precomputed context.
“a language task instruction through the dataset task field, or precomputed `context` and `context_mask` tensors”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/fastwam.mdx - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Policy type string is `fastwam`; default action_horizon 32.
“@PreTrainedConfig.register_subclass("fastwam")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/policies/fastwam/configuration_fastwam.py - T1 · upstream documentedrelease-notessource 2026-07-06verified 2026-10-24 (1d ago)lerobot 0.6.0
FastWAM was added in LeRobot v0.6.0.
“* feat(policies): Add FastWAM Policy by @Maximellerbach in https://github.com/huggingface/lerobot/pull/3834”
https://github.com/huggingface/lerobot/releases/tag/v0.6.0 - T1 · upstream documentedpapersource 2026-03-17verified 2026-10-24 (1d ago)
Fast-WAM paper states a 1B action expert and 6B total parameters on a Wan2.2-5B backbone.
“resulting in a 1B action expert and a total model size of 6B parameters.”
https://arxiv.org/abs/2603.16666v1 - T1 · upstream documentedpapersource 2026-03-17verified 2026-10-24 (1d ago)
Fast-WAM uses a joint flow-matching objective over action tokens and future video latents.
“Fast-WAM is trained with a joint flow matching objective over action tokens and future video latents.”
https://arxiv.org/abs/2603.16666v1 - T3 · community reportedhub-checkpointverified 2026-10-24 (1d ago)
Hub API: lerobot/fastwam_base (license:apache-2.0), lerobot/fastwam_libero_uncond_2cam224, lerobot/fastwam_robotwin_uncond_3cam_384, ZibinDong/fastwam_libero_uncond_2cam224 and yuanty/fastwam exist (no license tags on the latter four).
https://huggingface.co/lerobot/fastwam_base