Tasks

Ten tasks from eight datasets, in six families. Every task reads the same input, one wrist's accelerometer and gyroscope in a forward–left–up frame, and is defined by a task card that fixes the window, the decoder, the training recipe, the split, the selection rule and the scorer (see Task cards). The figures below show, for each family, the input window, what the encoder and the card's decoder produce, and how it is scored. Input traces are real Nymeria right-wrist windows; labels, captions and class scores are illustrative.

Pose estimation

	Arm pose arm_pose_nymeria	Full-body pose full_body_pose_nymeria
Dataset	Nymeria: real right-wrist IMU with Project-Aria motion capture, 300 h released	same
Input	10.24 s window at 50 Hz	same
Output	6-D rotations of the five right-arm joints (collar, shoulder, elbow, wrist, hand) per 0.32 s patch	6-D rotations of the 23 SMPL body joints per 0.32 s patch; the pelvis orientation comes from ground truth
Decoder	width-192 attention decoder, two cross-channel and two temporal layers	width-192 attention decoder, one cross-channel layer
Metric	forward-kinematics position error of the five joints, cm ↓	root-relative MPJPE over the 23 joints, mm ↓
Split	by recording session (867 training, 101 evaluation sessions); participants recur	same

Frozen features trail training from scratch on full-body pose; fine-tuning InrtlWrist reaches 88.1 mm against 92.7 mm from scratch.

Inertial odometry

	Inertial odometry odometry_nymeria_local
Dataset	Nymeria (same recordings as the pose tasks)
Input	2.56 s window at 50 Hz
Output	pelvis velocity in the wrist frame, four values per 0.32 s patch
Decoder	width-192 attention decoder, one temporal layer
Scoring	the predicted velocities are rotated into the world frame with the ground-truth initial wrist orientation propagated by integrating the gyroscope, summed into a trajectory and compared with the pelvis trajectory: absolute trajectory error, m ↓
Split	by recording session

The protocol uses the true initial orientation and short windows, so it is easier than full inertial navigation and its numbers are not comparable to world-frame odometry systems. Fine-tuned InrtlWrist reaches 0.154 m ATE against 0.162 m from scratch.

Microgesture recognition

	Microgesture microgesture_openwatch
Dataset	OpenWatch: five hand gestures (double clench, double pinch, pinch down, pinch up, slide) plus background, real recordings
Input	1 s windows at 100 Hz, six windows per clip from a centred 2 s interval
Output	class logits per window; a clip is labelled by the first run of three identical window predictions, else by majority
Decoder	linear head
Metric	clip macro-F1 on the test split ↑
Split	by participant over the 43 canonical identities (the release's 50 ids contain duplicated and mirrored recordings), stratified by posture

Frozen InrtlWrist reaches 64.7 clip F1 and stays ahead of every fine-tuned matched method on this task.

IMU–text retrieval

	IMU–text retrieval language_humanml3d
Dataset	HumanML3D motions with right-wrist IMU simulated from the motion capture (UniMTS release), with captions
Input	10.24 s window at 50 Hz
Output	an embedding aligned with the CLIP ViT-B/32 text space
Decoder	a projection, trained with paired motion–text supervision for every encoder alike
Metric	mean Recall@1 over IMU→text and text→IMU in galleries of 100 (200 random galleries on the official test split; the full gallery holds 3,701 clips) ↑
Split	the official HumanML3D split; 437 of the test clips share source frames with training clips, which affects every encoder equally

All encoders receive paired supervision during downstream training, so the task measures how well a pretrained representation supports learning alignment, not zero-shot alignment. Under frozen probing no supervised source task approaches the pretrained encoder here (27.9 Recall@1).

Parkinson's disease classification

	Parkinson's disease pd_pads
Dataset	PADS (PhysioNet): 469 participants, healthy control, Parkinson's disease or differential diagnosis; eleven scripted assessment steps recorded on both wrists
Input	2.56 s windows at 50 Hz; each (participant, step, wrist) recording is its own single-wrist session
Output	class logits per window; a participant's prediction is the argmax of the mean posterior over all of that participant's windows
Decoder	linear head, inverse-frequency class-weighted cross-entropy
Metric	participant macro-F1 over the three classes ↑ (balanced accuracy is reported too)
Split	subject-disjoint 80/20 (375 / 94 participants)

The participant metric is coarse: a never-trained frozen encoder reaches 51.7, so encoders within a few points of each other are not separated by it.

Activity recognition

	HHAR har_hhar	MHEALTH har_mhealth	PrISM har_prism	TransfHAR har_userwatch
Classes	6	12	5 global daily-living procedures	60 everyday hand activities
Subjects	9	10	31	10
Split	subject-disjoint, 20 % held out	subject-disjoint, 20 % held out	subject-disjoint, 20 % held out	80/20 in time within each recording, since most activities were performed by one participant

All four use 2.56 s windows at 50 Hz, a linear head and window macro-F1 ↑; the leaderboard's activity column is the mean over the four sets. These sets are held out of the released pretraining corpus; the activity datasets that are in the corpus (WISDM, PAMAP2, RecoFit, SAMoSA, UT-Watch) appear in the website's per-task table for reference only.