Repository navigation
Releases
Extended version (arXiv v1)
Code, results and paper source behind the extended version of Unreasonable Confidence from Simulator Misspecification: Directional Robustness on the SPEED+ Sim-to-Real Benchmark, as submitted to arXiv. The workshop version is the sim2science-2026-camera-ready release.
New since the camera-ready:
- paper/arxiv/: the LaTeX source, the PDF, make_tables.py, which writes every table from results/, and check_numbers.py, which checks every number in the text against results/ (83 claims, 0 failures).
- results/arxiv/: the extended version's analyses: rotation-error shape, coverage by range bin, and the synthetic-against-real transfer points.
- Weights: all 17 trained checkpoints are on Zenodo with a model card, under CC BY 4.0: https://doi.org/10.5281/zenodo.24140385. The scripts accept their published names (
<arm>.pt) as well as the names training gave them.
Unchanged, and still here:
- results/: every number in both papers as CSV, including the post-review experiments.
- results/predictions/: per-image predicted poses for the 20 runs behind the tables, with the SPEED+ labels (CC BY 4.0).
- preregistration/: the predictions and decision rules, committed before the results they predict were opened.
Every release of this repository is archived on Zenodo under https://doi.org/10.5281/zenodo.24011140. SPEED+ itself is not redistributed: download it from Zenodo record 5588480.
Sim2Science 2026 camera-ready
Code, results and paper source behind Unreasonable Confidence from Simulator Misspecification: Directional Robustness on the SPEED+ Sim-to-Real Benchmark (Sim2Science workshop, NeurIPS 2026), as submitted on 30 Oct 2026.
- results/: every number in the paper as CSV, including the post-review experiments.
- results/predictions/: per-image predicted poses for the 20 runs behind the tables, with the SPEED+ labels (CC BY 4.0).
- paper/sim2science-2026/: the LaTeX source, the PDF, and check_numbers.py, which checks every number in the text against results/.
- preregistration/: the predictions and decision rules, committed before the results they predict were opened.
- results/checkpoints.json: the sha256 of all 17 trained checkpoints. The weights are archived separately (link to follow).
SPEED+ itself is not redistributed: download it from Zenodo record 5588480.