--- license: other tags: - diffusion - control - controller-synthesis --- # DiffusionControl checkpoints Trained models for the DiffusionControl project (conditional diffusion models that generate certified feedback controllers). Code, protocols and results live in the GitHub repo `matCercola18/matte_server` under `DiffusionControl/`, branch `icml-experiments`; each family has a research doc in `DiffusionControl/docs/`. Files keep their repo paths (`models/...`), so code that loads `models/.pth` works after `huggingface_hub.snapshot_download`. Only final checkpoints are included (no `Last_*`, no intermediate step checkpoints). | family | files | doc | |---|---|---| | Furuta (QUBE-Servo 2), headline 1D U-Net generator | `models/v4/unet_a*_s96000.pth`, `unet_s_s40000.pth`, `unet_sxd_s40000.pth` | `furuta_generated_trajectory_v4.md` | | Furuta regression twins and MLP ablations | `models/v4/reg_*`, `models/v4/mlp_*` | same | | Furuta earlier trajectory models (v1-v3, clip-bug fix) | `Best_furuta_traj_cond_*` | `furuta_generated_trajectory*.md` | | Furuta energy-shaping models (superseded) | `Best_furuta_cond`, `_full_cond`, `_handoff_cond` | `diffusion_complete_nonlinear_controller.md` | | Pendulum | `Best_pendulum_*` (multi-solution `multi_*`, physics-informed `pg_*`/`abl_dl_*`, ablations `abl_*`) | `pendulum_multi_solution.md`, `physics_informed_generation.md`, `ablations.md`, `ood_shift.md` | | SISO Lur'e | `Best_nonlinear_cond_popov_*`, `siso_pg_c*`, `costcond_siso`, `cert_perfspan_L`, `regressor_siso` | RESULTS.md, `physics_informed_generation.md` | | MIMO 2x2 | `Best_mimo_cond_2x2*` | RESULTS.md | | CVAE baselines | `Best_cvae_*` | `ablations.md` | | Witsenhausen | `*witsenhausen_cond*` | `witsenhausen.md` | Note: `Best_nonlinear_cond_popov_L_sector/_closedloop/_distill` are documented negative results. ## Round 2 (2026-10-03): `models/round2/`, `data/round2/` Files here do not keep repo paths; point the loaders at them explicitly. | family | checkpoints | data | doc | |---|---|---|---| | Furuta, friction-conditioned (mp, Lp, Dr, Dp, Coulomb Cr, Cp) | `round2/furuta_v5/v5_s{0,1,2}_s96000.pth` | `data/round2/furuta/furuta_nmpc_bank_v5merged.npz`, `furuta_traj_dataset_v5.pt` | `furuta_friction_hardware_ready.md` | | Furuta, voltage-headroom (Vref) conditioned | `round2/furuta_f2/f2_s*_s8000.pth` | `furuta_bank_vref_*.npz`, `furuta_traj_dataset_f2.pt` | `furuta_robustness.md` | | Furuta, unified hardware model (friction + Vref, 7-d) | `round2/furuta_f3/f3_s{0,1,2}_s24000.pth` | `furuta_bank_f3.npz`, `furuta_traj_dataset_f3.pt` | `furuta_robustness.md` | | Unstable coupled MIMO, full-state (`mlp_*`, `asinh_*`) and output-feedback BMI (`sof*`, `c3*`) | `round2/mimo_unstable/` (`*_diff_*` diffusion, `*_reg_*` regression twin) | `data/round2/mimo_unstable/` | `mimo_unstable_family.md` | | Output-feedback scaling m=2/3/4, banks 1k/3k/8k | `round2/mimo_unstable/d2m{m}n{N}_*` | `data/round2/mimo_sof_scaling/` (banks, held-out sets, datasets, evaluated candidates) | `mimo_unstable_family.md` §D2 | | Output-feedback scaling m=2..5, final full recipe (96k x 64, 3 seeds), banks 8k (+16k for m=4) | `round2/mimo_sof_scaling_full/d3f{m}{n}_{diff,reg}_s*_s96000.pth` | `data/round2/mimo_sof_scaling/` (`d3f_m*.json` results) | `mimo_unstable_family.md` §D3 | | Furuta online re-synthesis on the unified F3 model (episode logs) | — | `data/round2/furuta_resynth_b2/` | `furuta_online_resynthesis.md` §B2 | | SOF scaling, batch-256 equal-sample recipe (equivalence check for D3) | `round2/mimo_sof_scaling_b256/d3m*_s24000.pth` | `data/round2/mimo_sof_scaling/` | `mimo_unstable_family.md` §D3 | ## QUBE-Servo 2 hardware (2026-10-08): `models/round2/qube_*`, `data/round2/qube_*` Fine-tuned on the measured lab unit (arm inertia, motor gain and cable spring identified on the rig; base JSON `qube_params_rig1.json`). Hardware report: `DiffusionControl/docs/qube_hardware_report.md`. | model | what | hardware | |---|---|---| | `qube_rig1/rig1_s6000.pth` | F3 fine-tuned on a 2000-plant bank around the unit (mp, Lp +-10%) | 16/16 caught (computed TVLQR tracking) | | `qube_rig1w/rig1w_s6000.pth` | same, mp and Lp +-30% (added-mass / other pendulum) | not yet run | | `qube_rig1wg/rig1wg_s8000.pth` | fully generated controller: also outputs the TVLQR gain schedule K(t) (5-channel U-Net) | 6/6 caught | | `qube_armprov/armprov_s6000.pth` | provisional (pre-sysid) arm, superseded | - |