Abstract—Dynamic-0 (X0) addresses model-free 6D pose estimation for unseen objects—without Computer-Aided Design (CAD) models—while remaining efficient across varied hardware. It uses a dual-agent design: the first agent achieves coarse pose alignment via DINOv3 visual descriptors, Facebook Artificial Intelligence Similarity Search (FAISS)-based retrieval, and Efficient Perspective-n-Point (EPnP) with Super Random Sample Consensus (SuperRANSAC); the second selectively refines this estimate using differentiable 3D Gaussian Splatting (3DGS) rendering, improving geometric precision only where needed to avoid uniform computational overhead. To adapt inference dynamically, X0 integrates multi-scale routing, sparse feature reactivation, and a novelty-aware semantic gating mechanism based on generative density scoring over DINOv3 and MesaNet embeddings. A MesaNet temporal block further stabilizes pose sequences by treating temporal consistency as a locally optimal test-time training problem, solved via conjugate-gradient updates. Evaluated under the Benchmark for 6D Object Pose Estimation (BOP) Challenge 2025 model-free protocol on BOP-H3 datasets (HOPEv2, HANDAL, HOT3D), Dynamic-0 achieves an Average Precision of 0.473 on HOPEv2, with adaptive throughput of 2.27–15.4 Frames Per Second (FPS) and a 73% reduction in temporal jitter (average processing time: 40.98 s/image). Additional results are reported on HANDAL and HOT3D leaderboards, scoring as the first one in terms of accurate precision and average time on the leaderboards of all three datasets of the BOP Challenge 2025. Code is publicly released at the GitHub repository; the official leaderboards are available at the links indicated in the Data Availability section of this paper. These findings suggest that combining self-supervised representations, adaptive computation, generative novelty detection, and temporal optimization offers a practical route toward scalable, hardware-aware 6D pose estimation in real-world, model-free settings.
 
Keywords—6D pose estimation, model-free detection, zero-shot learning, adaptive inference, MesaNet, high-performance computing
 
Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).