multizone-v02/README.md

Multi-zone actuation · v0.2 supplement

4,080 equilibrium states in 680 independent control scenarios. This supplement adds 3, 6 and 12 actuation regions to one continuous, planar extensible rod. Eight-region scenarios are a held-out zone-count test. It does not represent connected robots, contact or physical-time dynamics.

Protocol

• 200 scenarios each at 3, 6 and 12 zones; 80 additional scenarios at 8 zones.
• Each scenario has six loading levels t=0,0.2,...,1; all six stay in the same split.
• Train: 2,520 states; validation: 540; test: 540; zone-count OOD: 480.
• Curvature controls sampled uniformly in [-2.4,2.4]; chi log-uniform in [1e-4,10^-2.5]. Final fx/chi in [0,0.6], fy/chi in [-0.6,0.6]. Curvature and forces ramp together with t. L=EA=1, zero end moment and zero microforce h.
• A 48-element angle-FE model generates locally stable equilibrium states. 34 final states are checked against 96 elements. Maximum checked endpoint difference: 4.604e-5 L.
• Zero-loading states are repeated across scenarios by construction; they are not evidence of generalization. Evaluation reports both all-state and nonzero-loading errors. Grouping prevents a nonzero loading path from crossing splits.
• Eight-zone OOD is a limited zone-count shift, not arbitrary-geometry or load extrapolation.

Each compressed JSONL record contains scene and split IDs, zone count, controls, loading parameter, angle DOFs, a 65-point reference-coordinate centerline, strain, endpoint, residual, local eigenvalue and Newton trace. Rejected samples are retained in rejected.json (none in this release). Full metrics and mesh checks accompany the data.

Reproduce with OPENBLAS_NUM_THREADS=1 python multizone.py from the original project layout. Source files are supplied in source/; they resolve the project root as their parent directory's parent. Dependencies: NumPy, SciPy, PyTorch and safetensors. Generation and training use seed 20261015.

中文

新增 680 组驱动工况、4,080 个平衡状态。3、6、12 分区用于训练与同分布测试；8 分区全部留作测试。同一工况的加载状态不跨数据划分。数据描述连续柔性杆的理想化曲率驱动，不是实物机器人实验数据。

Data: CC BY 4.0. Source code: Apache-2.0. Original v0.1 files remain unchanged.