dinobotpose

DINObotPose weights

Paper — DINObotPose: Monocular Robot Pose and Joint Angle Estimation with a Pretrained Vision Foundation Model and Kinematic Fitting (ICRA 2027 submission, under review) Code — Najongs/DINObotPose release v1.0.1: pinned dependencies, checkpoint SHA-256 manifest, an environment doctor.py, and a one-command reproduction of the main comparison table and the occlusion curve.

Monocular robot pose and joint-angle estimation from a single RGB image, with no encoder readings and no ground-truth bounding box. These are the released checkpoints for Najongs/DINObotPose; the code, the evaluation entry points, and the reproduction script live there.

Built with DINOv3

These weights contain the DINOv3 ViT-B/16 trunk and are a derivative of the DINOv3 Materials. They are distributed under the DINOv3 License Agreement, not under the MIT license that covers the code. Downloading them is acceptance of that agreement; redistributing them onward is permitted only under the same terms, with a copy of the agreement and the "Built with DINOv3" notice included. Publications reporting results obtained with them should acknowledge DINOv3 (Siméoni et al., 2025).

Layout

The trunk ships once rather than once per detector, with a small override for the detectors whose continue-training moved their last blocks: 741 MB downloaded against 1486 MB assembled, or 517 MB for Panda alone.

manifest.json               what to fetch and how to reassemble it
dinov3_vitb16_trunk.pth     the shared trunk, 211 tensors
panda/, kuka/               per-robot keypoint heads, trunk overrides, angle and rotation heads

Fetch and assemble with the repository's own script, which reproduces the six checkpoints per robot exactly:

python scripts/download_weights.py --robot panda
python scripts/pack_weights.py verify     # tensor-by-tensor proof that assembly round-trips

---

Jongyeol Na — Researcher, KIRO · M.S. DGIST, IROM Lab GitHub · Knowledge graph · W&B