Upload docs/run/spider.md with huggingface_hub

docs/run/spider.md

@@ -0,0 +1,91 @@
+# Spider — Run Commands
+
+## What this method does
+Sampling-based retargeter (no RL). Simulator: **MuJoCo-Warp** (Python 3.12, CUDA 12.1). Produces per-cell scene.xml + kinematic trajectory + optional MJWP-sampled rollout. **No training** — this is the fundamental reason Spider is fast.
+
+## 5-stage preprocess pipeline (per cell)
+
+1. `process_datasets.oakink` — MANO → per-object keypoints + task_info.json
+2. `preprocess.decompose_fast` — voxel-grid convex decomposition of object meshes
+3. `preprocess.generate_xml` — assemble scene.xml + scene_eq.xml
+4. `preprocess.ik_fast` — IK from MANO keypoints → trajectory_kinematic.npz (**CPU-only**)
+5. `run_mjwp.py` — MuJoCo-Warp sampling → trajectory_mjwp.npz (**GPU-required**)
+
+All 5 stages per cell via the canonical wrapper:
+
+```bash
+docker compose run --rm --entrypoint bash spider -c \
+    "bash /workspace/shared/scripts/spider_preprocess.sh \
+         /workspace/data/oakink_v2 <task> <hand> bimanual 0"
+```
+
+Where `<task>` is the traj name stripped of `_bimanual` (e.g. `lift_board`).
+
+## Run one eval cell
+
+```bash
+./scripts/run.sh spider \
+    --hand {allegro|inspire|schunk|xhand} \
+    --dataset oakink_v2 \
+    --traj <traj-id> \
+    --seed 0 \
+    --warmstart default \
+    --viz none --dev
+```
+
+Example:
+```bash
+./scripts/run.sh spider --hand schunk --dataset oakink_v2 \
+    --traj lift_board_bimanual --seed 0 --viz none --dev
+```
+
+## Batch preprocess + sample for ALL 28 cells
+
+```bash
+# idempotent; skips cells that already have trajectory_mjwp.npz
+nohup /tmp/spider_mjwp_batch.sh > /mnt/external/outputs/logs/spider_mjwp_batch/ALL.log 2>&1 &
+```
+
+Current state (as of last check):
+- kinematic: 28/28 ✅
+- mjwp: 28/28 ✅ (schunk + xhand finished; inspire gap batch in progress)
+
+Budget: ~30 min/cell average (wide range: 7 min for simple tasks, 78 min for pick_spoon_bowl). Total for 28 cells ≈ 7–10 hr if everything is from scratch.
+
+## Spider × Arctic (via patch, not default)
+
+Upstream has no arctic process_datasets. `methods/spider/patches/arctic_support.patch` adds `spider/process_datasets/arctic.py` + arctic hydra override — **partial**: keypoint extraction only, no scene-generation or mjwp arctic path yet. See `docs/46_gap_plan.md` Gap 3.
+
+## Regenerate standardized video for all Spider cells
+
+```bash
+MUJOCO_GL=egl /home/azureuser/ckwolfe/benchmarking/.venv-shared/bin/python3 \
+    scripts/render_standardized.py --method-label spider
+```
+
+Outputs: `outputs/videos_std/spider_<run_id>.mp4`, 720×480 @ 30fps, BENCH_CAMERA front view.
+
+The renderer prefers `trajectory_kinematic.npz` (uniform 2D qpos) over `trajectory_mjwp.npz` (receding-horizon planning tensor) for consistent rollout length across the grid.
+
+## No training scripts
+
+Spider is sampling-based; there is no `train.py` with `--train` flag beyond the MJWP sampling wrapper:
+
+```bash
+./methods/spider/scripts/train.py --hand <h> --traj <t>_bimanual \
+    --task full --train
+```
+
+`--task full` = preprocess + sample (stages 1-5). `--task preprocess` = stages 1-4 only (CPU-only). `--task sample` = stage 5 (GPU).
+
+## Status check
+
+```bash
+python3 scripts/sanity_check.py | grep "oakink processed"
+# expect:
+#   mano:    7/7 ✅
+#   allegro: kinematic=7/7, mjwp=7/7 ✅
+#   inspire: kinematic=7/7, mjwp=3/7 (or 7/7 if inspire batch finished)
+#   schunk:  kinematic=7/7, mjwp=7/7 ✅
+#   xhand:   kinematic=7/7, mjwp=7/7 ✅
+```