Learning humanoid state estimation and language-grounded motion understanding from robot recordings ties training coverage to the robot's motion repertoire. Human motion datasets offer broader coverage, yet transferring this knowledge requires overcoming differences in body geometry, sensing, and physical disturbances. To enable this transfer, we propose a unified inertial encoder whose shared inertial interface uses kinematic IMU synthesis to represent human and humanoid motion at corresponding body sites. Within this common input space, joint training on human and kinematically retargeted humanoid motion combines velocity estimation and motion--text alignment with continuous-time encoding and sensor and motion randomization. The resulting model transfers zero-shot to a physical Unitree G1 without physics simulation, real-robot training trajectories, or fine-tuning, achieving the lowest velocity error among evaluated methods across five datasets, including 0.099m/s on real G1. Its motion-description retrieval also extends to retargeted G1 motion categories observed only in human training, demonstrating how human motion knowledge can support the generalization of humanoid inertial odometry and language-grounded motion understanding.
Figures & tables
Fig. 2: Overview of the proposed method. Left: kinematic transport combines pelvis IMU measurements, joint encoders, and link kinematics to synthesize inertial signals at anatomically corresponding sites. Center: continuous convolution converts native-rate IMU streams into fixed-duration tokens. A shared transformer alternates within-site and full attention, with heads predicting body-frame velocity, gravity-free acceleration and clean angular velocity, and a motion embedding aligned with frozen text features. Right: example outputs on a physical G1 illustrate inertial odometry and motion-description retrieval.
Parameter
Setting
Motion and sampling
Gravity magnitude
U(9.76,9.83) m/s 2
Body scale
U(0.75,1.0) / U(1.0,1.33)
Motion speed
U(0.7,1.3) / U(0.9,1.6)
Human sampling rate
Target U(120,240) Hz
Timestamp jitter
Spacing CV target U(0,0.15)
TABLE I: Training augmentation for motion diversity, sampling variation, and hardware disturbances.
Dataset
Emb.
Motion
Real IMU
Freq.
Hours
Lang.
Train
Nymeria
Human
Real
✔
240
300+
✔
✔
Natural Motion
Human
Real
✔
240
40.6
✗
✔
BONES-Seed
Human
Real
✔
120
144
✔
✔
BONES-Seed
G1
Retarget
✔
120
144
✔
✔
Real-Trajectory
G1
Real
✗
200
0.7
✗
✗
TABLE II: Datasets used in this work. Real G1 recordings are reserved for evaluation. Emb. denotes embodiment; Freq. is in Hz; Lang. and Train mark language supervision and training use.
Human
Humanoid (G1)
Nymeria
Natural Motion
BONES-Seed (Human)
BONES-Seed (G1)
Real-Trajectory
Method
AVE
ATE GT
ATE Pred
AVE
ATE GT
ATE Pred
AVE
ATE GT
ATE Pred
AVE
ATE GT
ATE Pred
AVE
ATE GT
ATE Pred
Pelvis IMU only
TLIO † [ 9 ]
0.122
8.149
17.76
0.189
40.72
81.31
0.245
0.717
0.719
0.173
0.761
0.762
0.371
17.86
17.86
AirIMU † [ 15 ]
12.62
–
–
109.3
–
–
0.212
9.472
505.5
0.281
11.83
306.6
6.231
731.1
–
AirIO † [ 10 ]
0.069
7.481
113.9
0.280
158.1
157.9
0.140
0.530
1.918
0.118
0.720
1.690
0.350
16.82
20.48
TABLE III: Inertial odometry across human and humanoid datasets. AVE measures velocity error (m/s); ATE measures trajectory error (m), using ground-truth (GT) or gyro-integrated (Pred) attitude. Lower is better; bold marks column minima.
Fig. 3: Estimated trajectories sampled from each dataset. Predicted pelvis velocities from our model (gold) and AirIO (blue) are integrated using ground-truth attitude; gray shows the reference trajectory. Each panel reports per-recording ATEGT in meters (Ours/AirIO).
Human
Humanoid (G1)
Method
Nymeria
BONES-Seed
Seen cat.
Unseen cat.
Collapse
4.4
/
0.1
19.7
/
0.2
21.9
/
0.8
9.8
/
0.1
ImageBind [ 20 ]
8.9
/
0.1
7.1
/
0.4
11.4
/
1.6
4.6
/
0.1
IMU2CLIP † [ 11 ]
16.2
/
0.4
51.8
/
5.2
36.1
/
11.2
17.3
/
0.5
Ours
14.9
/
0.6
60.7
/
15.9
63.4
/
32.9
53.8
/
11.3
TABLE IV: Motion caption retrieval: motion-to-text recall at 1 (R@1, ↑ ) / text-to-motion mean average precision (mAP %, ↑ ). Unseen G1 categories occur only on human embodiment in training dataset, showing the cross-embodiment knowledge transfer.
Fig. 4: Motion-description retrieval on a retargeted G1 sequence. Sampled poses (top) accompany annotated categories and descriptions retrieved by our model and ImageBind (bottom). Our model uses 2s and ImageBind uses 10s windows; hatching marks intervals without a complete input window.
Fig. 5: Inference runtime and deployment on a physical G1. Left: latency on an NVIDIA L40S with 200-Hz inputs; labels show IMU count, window duration, and model size. TRT denotes TensorRT. Right: stair descent shown by robot poses over a 3D point cloud, with retrieved descriptions changing from “Walking” to “Standing” as the robot stops.
Nymeria
BONES-Seed (G1)
Real-Trajectory
Variant
ATE GT
ATE Pred
ATE GT
ATE Pred
ATE GT
ATE Pred
1. Training embodiment
Humanoid-only
8.041
17.50
0.162
0.180
5.901
13.15
2. Input representation
Single pelvis IMU
3.592
16.66
0.330
0.335
8.114
12.81
Pelvis IMU + joints
2.283
16.12
0.136
0.164
3.513
18.36
TABLE V: Model and training recipe ablations (ATE in m, lower is better), grouped by the four questions in Sec. IV-E .
Fig. 6: Human and humanoid motion in the learned feature space. A t-SNE projection of [CLS] embeddings uses blue for humans, gold for G1, and marker shapes for datasets. Dashed regions enclose 60% of windows in each motion category, showing overlap between human and humanoid data.
Fig. 7: Model capacity and training exposure under a fixed compute budget. (a) Measured velocity error and fitted compute-optimal trend (dashed). (b) Fitted error surface with observations (blue), constant- 6ND contours (gray), and predicted optima (gold). N counts parameters and D counts processed tokens, including repeats; the surface includes extrapolation.
The University of New South Wales · The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology