FAQs

Registration and teams

• Do I need to register if I am competing alone?

Yes. Solo participants register as a team of one. No submission is evaluated without a prior registration.

• Can I join after the registration deadline?

No. Registration closes on October 27, 2026 (extended once), and there are no extensions.

• Is there a team size limit?

No, but every member must contribute and appear as an author on the paper.

• Can I be on more than one team?

No. Individuals belong to one team. Supervisors and advisors may be associated with multiple teams, provided the teams do not share code or results.

• Can two teams from my lab both enter?

Yes, if they work independently. Two teams sharing code, weights, or curated data are one team, and will be scored as one.

• Do I have to attend the conference?

At least one team member must present the paper. Contact the organisers early if this presents a difficulty.

Data

• Where do I get the data?

From UC Berkeley, directly. BOXRR-23 is public. The organisers distribute no data and no code.

• Do I have to download all 8 TB?

No, and you should not try. The sweep and velocity filters on the Data page exist precisely so that you download only the sessions worth training on. Apply the filter during download, discarding recordings as you go.

• How much data will survive the filters?

That depends on your implementation choices. Expect the qualifying fraction to be a small minority of the dataset — that is the intent.

• Why is the session filter defined on Euler angles when scoring uses quaternion geodesics?

They are answering different questions. The filter asks whether the head moved persistently along any single rotational axis, which is most naturally expressed per axis. Scoring asks how far a predicted orientation is from the true one, which is a geodesic question. Use the definitions as given, and do not validate your model on the filtering measures.

• Can I use additional datasets?

Yes. Any data you are legally entitled to use is permitted, provided it is disclosed in your paper. The evaluation set is unreleased and drawn from outside BOXRR-23, so broader training data may well help.

• Can I use a pretrained model?

Yes, if it is publicly available and properly cited. Models pretrained on the unreleased test set do not exist, so there is nothing to guard against here.

• My sessions have irregular timestamps. What should I do?

Resample to a uniform 100 Hz grid using spherical linear interpolation. The evaluation stream is uniform, and the task is defined in samples. Training on irregular data means training under a different regime from the one you are scored in.

• Should I train on sessions outside 85 to 110 Hz?

The specification says to filter them out, and there is good reason: a fixed 20-sample horizon is only a fixed 100 millisecond horizon when the clock is fixed. Sessions at other rates make your validation numbers mean something different from your score.

The task

• Why 400 samples of context?

It is an upper bound, not a target. Four seconds is the most you may use. Shorter contexts are permitted and are rewarded through the latency criterion.

• Must I use all 400 samples?

No. Accept the full input tensor and slice internally if you use fewer.

• Is the model scored on all 20 predicted samples, or only the 20th?

Accuracy is measured on the 20th sample only. The other 19 are still required in the output, and predicting the full trajectory coherently generally helps the 20th.

• Why the 20th sample and not an average over the horizon?

Because it is the hardest point and the one that matters for the target applications. A model that excels at 10 milliseconds and fails at 100 is of little use for predictive rendering.

• Can my model output a distribution, or multiple hypotheses?

The submitted graph must emit a single [1, 20, 4] tensor. How you arrive at it — including by collapsing a distribution to its mode or mean — is your choice, and worth describing in your paper.

• Is the model allowed to see its own past predictions?

Yes. Autoregressive decoding within the 20-step horizon is fine. What is forbidden is seeing ground-truth samples at or beyond the prediction boundary.

• What does "streaming" mean in practice for a stateless model?

The harness feeds the shifted 400-sample window and re-runs the graph. Nothing else changes. Stateless models are entirely acceptable; they simply do more arithmetic per step, which the latency criterion reflects.

Evaluation

• What exactly is the error metric?

The geodesic angle between predicted and ground-truth orientation at the 20th sample, averaged over all evaluation windows, in degrees. It is the mean, not the RMS, and not one minus the absolute dot product. These differ substantially on the same model — RMS is typically around twice the mean on this kind of data — so validate against the right one.

• Will I see my score before the deadline?

No. There is no leaderboard and no intermediate feedback. The test set is unreleased, and submissions are evaluated once, after the deadline.

• Can I submit more than once?

One submission per team is evaluated. If you send a corrected archive before the deadline, the most recent one is used.

• How is "innovation" judged? It sounds subjective.

It is a judged criterion, ranked by the organising committee across all submissions, and it is 20 percent of the score. It rewards genuinely new architectural ideas, compression strategies, or training techniques — and it rewards them most when the paper explains them clearly enough to be reproduced. A strong result from a standard architecture applied competently will rank mid-field here, not last.

• Which hardware is latency measured on?

An XR-class edge-AI target of the kind described on The Challenge page. Participants need no hardware of their own. Because you cannot profile the target, the operator constraints on the Submitting Artifacts page are the actionable guidance — follow them.

• If nobody can measure latency, how do I optimise for it?

Report MACs per streaming step, keep your graph inside the safe operator set, and prefer stateful architectures that avoid recomputing the full window on every sample. Those three things account for most of the achievable difference.

• Why is accuracy only 40 percent of the score?

Because a head-pose predictor that cannot run on a headset solves nothing. The weighting reflects deployability as a first-class requirement rather than an afterthought.

• What happens if two teams tie?

Tied criteria receive the average of the tied ranks. A tie in the final score is broken by the accuracy criterion.

Submission

• Which quaternion order does the harness use?

Scalar first: (w, x, y, z), for both input and output. If your pipeline uses scalar-last, reorder at the model boundary before export.

• Must I submit ONNX?

PyTorch and TensorFlow are accepted, but ONNX is what gets deployed. A model that will not export cleanly is at a real disadvantage on latency.

• My architecture needs an operator on the "avoid" list. What now?

Contact the organisers before the deadline. Explain what you need and why. It may still be deployable, and it is better to find out in October than in November.

• Can I submit both a float and a quantised model?

Yes. Include both, and state clearly in your model card which is your primary submission.

• Is there a starter kit or baseline model?

No. The pipeline and the model are both yours to build.

• What if my submission fails a validity check?

You will be notified once and given 72 hours to fix the packaging. This covers mechanical faults only, not model changes.

Paper

• Three pages including references?

Yes. IEEE two-column conference format, references counted within the three pages.

• Can I submit the paper without the model, or the model without the paper?

No. A submission requires both.

• What must the paper report?

At minimum: your method, your own evaluation results including a comparison against a constant-velocity baseline, your total trainable parameter count, your MACs per streaming step, and a citation of BOXRR-23.

• Will papers be published?

Yes, with the ABC 2027 proceedings.

Anything else

My question is not here.

Email swapnilsayan.saha@st.com and swapnilsayansaha@ieee.org. If it is of general interest, the answer will appear on this page.