LimX Dynamics †Project Lead *Corresponding Author
A pose is a rotation, and a squared error on rotations cannot tell motion that flows from motion that shakes. FMG supervises how rotation changes over time rather than only where it lands, which is what makes a generated motion smooth enough to watch. Because a rotation retargets onto any skeleton, the same output also runs on a real humanoid, at any length asked for.
Text-to-motion still inverts a diffusion process and has not inherited the stack that now produces images: a single-stream DiT with joint attention over text and motion tokens, trained with a rectified flow. That transplant is not a drop-in, and the reason is geometric.
The target is a sequence of joint rotations, so a squared error in a rotation embedding is blind to whether error is spread smoothly or concentrated in frame-to-frame reversals, which few solver steps no longer average away; smoothness falls below the diffusion baseline it was meant to improve on. The missing supervision is the temporal derivative of rotation. FMG predicts the clean sample rather than a velocity field, which keeps every term of the loss a decodable pose, and supervises rotational velocity and acceleration; smoothness returns past the diffusion baseline at eight solver steps.
We assemble the 1,983 hours this needs, most of it our own capture, and score six axes against levels measured on that capture rather than idealised targets, so smoothness cannot reward frozen motion nor physical plausibility punish a body for contacting itself. Because the output is a rotation per joint, a sample retargets onto a robot's own skeleton: one set of outputs runs on two different physical humanoids.
A velocity prediction is a quantity in noise space that cannot be decoded. Predicting the clean sample keeps every term of the loss a pose, convertible to rotation matrices and pushable through forward kinematics.
A squared error in 6D rotation space is indifferent to whether error is spread smoothly or concentrated in frame-to-frame reversals. We supervise the body-frame angular increment and its difference, which is the quantity the metric is blind to.
Sampling tiles 240-frame windows with 60-frame overlap and blends them inside every denoising step, so a seam introduced at one step is denoised away by the next. Cost grows linearly and no context length is ever exceeded.
A pose given as a rotation per joint is defined on any skeleton with those joints, so reaching a machine is a retarget rather than a regeneration. One set of outputs ran on two different humanoids.
Each axis is 0–100 against fixed constants, so one run is comparable to another. Text-motion match, naturalness and diversity are scored in a contrastive text–motion embedding. Smoothness is the jerk of the body-frame angular increment, and physical plausibility is limb-capsule interpenetration after forward kinematics; both are referenced to the level real capture registers rather than to zero. Speed is per-request latency at batch size one.
| Method | Institute | Params | Nat. | Match | Div. | Smth. | Speed | Phys. | Avg |
|---|---|---|---|---|---|---|---|---|---|
| MotionMillion | Shanghai AI Lab (2025) | 7.8B | 42.3 | 30.5 | 51.7 | 49.5 | 45.5 | 72.1 | 48.6 |
| HY-Motion | Tencent Hunyuan (2025) | 1.0B | 38.0 | 52.6 | 67.3 | 54.3 | 31.4 | 71.4 | 52.5 |
| Kimodo | NVIDIA (2026) | 283M | 36.8 | 51.0 | 59.1 | 79.0 | 49.5 | 93.4 | 61.5 |
| GENMO | NVIDIA (2025) | 523M | 44.2 | 28.0 | 52.3 | 79.8 | 20.6 | 68.1 | 48.8 |
| FMG | LimX Dynamics (2026) | 495M | 86.4 | 86.8 | 89.3 | 90.7 | 81.4 | 85.0 | 86.6 |
The same prompt through five models. Every row of a prompt shares one camera and one crop window, so a body that sits lower or reaches further really has.
Nothing in the recipe fixes a length.
Training sees 8-second windows. Sampling tiles any requested length with overlapping windows and blends them inside every denoising step, so a seam raised at one step is denoised away by the next.
Across several hundred seams, the jerk where two windows meet is not distinguishable from no seam at all. Cross-fading finished windows instead only moves the kink to the edge of the fade.
The overlay reports the active prompt and the tiling boundaries.
Because the output is a rotation per joint, a sample retargets onto a robot's own skeleton. One set of outputs was executed on two different machines, both tracking the retargeted motion under the same whole-body controller.
One checkpoint, one sampling budget, no per-prompt tuning. Drag or scroll sideways; click any clip to enlarge.
@article{fmg2026,
title = {{FMG}: A Foundation Motion Generator for Text-Driven Humanoid Motion},
author = {Wang, Zhao and Hu, Jiangtao and Yu, Jack and Yu, Tao},
journal = {arXiv preprint arXiv:XXXX.XXXXX},
year = {2026}
}