The Challenge

Challenge goal

Develop a foundation model for head-pose prediction that:

• Accurately predicts n samples of future head pose given a window of context quaternion samples.

• Generalises to any user zero-shot, without fine-tuning.

• Is deployable on commercial XR edge-AI platforms.

Head motion is smooth over short horizons but sharply non-stationary over long ones. A model that merely extrapolates velocity will do respectably at 10 ms and poorly at 100 ms. Closing that gap is the substance of this challenge.

Task specification

• Input: Normalised head-motion quaternions, 4-D, tagged 100 Hz

• Context window: At most 400 samples (4 seconds). Shorter windows are permitted and are rewarded through the latency criterion.

• Prediction horizon: Exactly 20 samples (100 ms of future head pose)

• Causality: Strictly causal. The model may not access any sample at or beyond the prediction boundary.

• Fine-tuning: Not permitted on the test set, nor on any per-user data at inference time.

Primary metric

Accuracy is measured as the geodesic angular error at the 20th predicted sample — the true angle between the predicted and ground-truth rotation at 20th sample:

The absolute value of the inner product handles quaternion double cover, since q and −q represent the same rotation.

The reported figure is the mean of θ over all evaluation windows, expressed in degrees.

Please read this carefully when building your own validation. The metric is the mean angle at 20th sample over all evaluation windows — not RMS, and not 1−∣⟨q^,q⟩∣ Head-motion error distributions are heavy-tailed, and these three quantities differ substantially on the very same model: an RMS figure is typically around twice the mean. Teams that validate against a different statistic will systematically misjudge their standing. 

Evaluation criteria

• Accuracy: 40%, mean geodesic error at sample 20, on the unreleased test set

• Parameter count: 10%, total trainable parameters, verified against the delivered artefact

• Latency: 30%, measured by ST on an XR edge-AI target: batch size 1, steady state, median of repeated runs after warm-up

• Innovation: 20%, novelty of architecture, compression, or training technique, ranked by the organising committee across all submissions

The four criteria are measured in incompatible units — degrees, millions of parameters, milliseconds, and a qualitative judgement. They are therefore converted to a common scale by rank normalisation before being combined. For each criterion, submissions are ranked from best to worst and mapped onto the interval [0,1]:

where rank 1 is the best submission and N is the number of valid submissions. Tied submissions receive the average of the tied ranks. The final score is:

The lowest final score wins. A submission that ranked first on all four criteria would score 0.

Rank normalisation is used rather than raw values because a single outlier — one very large model, or one unusually slow one — would otherwise compress all remaining submissions into an indistinguishable band. It also means absolute test-set figures need never be published, which keeps the hidden test set hidden.

The test set

STMicroelectronics will evaluate all submissions on an unreleased test set that is not part of the BOXRR-23 dataset. It is not distributed to participants at any point, before or after the challenge.

Many challenges release their test set and rely on good faith. This one does not need to. The only route to a good score here is a model that genuinely generalises to people it has never seen, recorded on hardware it has never encountered — which is precisely the property a foundation model is supposed to have.

Designing for the edge

Criterion 3 carries 30% of the score, so efficiency is not an afterthought. Techniques we expect to see, and encourage:

• Quantisation — post-training quantisation or quantisation-aware training

• Pruning — structured pruning maps to hardware far better than unstructured

• Knowledge distillation from a larger teacher model

• Hardware-aware neural architecture search

• Efficient sequence architectures — temporal convolutional networks, state-space models, linear attention

• Operator-set reduction for on-device execution compatiblity

Participants are not expected to own XR hardware. ST performs all latency measurement. What you are expected to do is keep your model inside a portable, on-device friendly operator set.

What you deliver

• A trained model in ONNX, PyTorch, or TensorFlow format, with documentation for loading it and running inference on a standard Linux machine with CUDA support.

• Either 4-6 pages in IEEE format or 8-10 pages IJABC format scientific paper describing your work.

• Presentation by at least one team member at the conference.

The 1st deliverable should be shared as a Google drive link accessible to anyone. The paper deliverable sharing method will be discussed later.

Contact

Swapnil Sayan Saha, Software Design and Development Staff Engineer

STMicroelectronics Inc.,

swapnilsayan.saha@st.com AND swapnilsayansaha@ieee.org (make sure your query goes to both emails)