Upload docs/run/ours.md with huggingface_hub

docs/run/ours.md

@@ -0,0 +1,73 @@
+# Oracle (ours) — Run Commands
+
+## What this method does
+Kinematic replay of Tara's retargeter output. Simulator: **MuJoCo** (Python 3.10, no CUDA required). **No training in this container** — Tara's retargeter is trained externally; this consumes the `.npz` output. Paper column = open-loop upper bound on trajectory quality.
+
+## Run one eval cell
+
+```bash
+./scripts/run.sh ours \
+    --hand {allegro|inspire|schunk|xhand} \
+    --dataset oakink_v2 \
+    --traj <traj-id> \
+    --seed 0 \
+    --warmstart ours \
+    --viz none --dev
+```
+
+Example:
+```bash
+./scripts/run.sh ours --hand allegro --dataset oakink_v2 \
+    --traj lift_board_bimanual --seed 0 --warmstart ours --viz none --dev
+```
+
+## Regenerate standardized video for all 28 Oracle cells
+
+**This is the canonical path** — `scripts/render_standardized.py` IS the Oracle rendering pipeline. It takes each cell's `trajectory_kinematic.npz` (or mjwp/ikrollout when higher-fidelity), the Spider-preprocessed scene.xml, and renders through MuJoCo with BENCH_CAMERA.
+
+```bash
+MUJOCO_GL=egl /home/azureuser/ckwolfe/benchmarking/.venv-shared/bin/python3 \
+    scripts/render_standardized.py
+```
+
+Produces all 28 cells in ~2–3 min total. Output:
+```
+/mnt/external/outputs/videos_std/std_<hand>_oakink_v2_<traj>_bimanual_seed0_default.mp4
+```
+
+All at 720×480, 30 fps, h264 crf 20, yuv420p. Frame counts match the kinematic trajectory length (100–1740 frames per cell).
+
+Per-cell re-render:
+```bash
+MUJOCO_GL=egl /home/azureuser/ckwolfe/benchmarking/.venv-shared/bin/python3 \
+    scripts/render_standardized.py --hand allegro --traj lift_board --force
+```
+
+## No training (this is an Oracle)
+
+`methods/ours/scripts/train_stub.py` is a no-op that records "training lives outside this repo" in a stub MetricsRow. There is no RL. The "method" is Tara's retargeter which runs in a separate codebase; we only consume its .npz output.
+
+## Current blocker — Tara's WarmStartTraj payloads
+
+```bash
+# Verify payload state
+/home/azureuser/ckwolfe/benchmarking/.venv-shared/bin/python3 -c "
+import numpy as np
+p = 'data/warmstart/ours/allegro/oakink_v2/lift_board_bimanual.npz'
+z = np.load(p, allow_pickle=True)
+print({k: z[k].shape for k in z.files if hasattr(z[k],'shape')})
+# expect: {'qpos': (0, 1), 'wrist_pose': (0, 7), 'object_pose': (0, 1, 7)}
+# qpos.shape[0]==0 → stub placeholder
+"
+```
+
+55/56 cells are currently T=0 placeholder stubs (`__meta__`: "DO NOT TOUCH — placeholder for Tara's retargeter output"). Only `xhand × oakink × lift_board_bimanual` has a real payload.
+
+`scripts/render_standardized.py` sidesteps this blocker by reading `trajectory_kinematic.npz` (Spider's IK output, always present) as the fallback. This gives 28 cells of "what the reference trajectory looks like under the shared MuJoCo renderer" — which is the honest Oracle definition.
+
+## Status check
+
+```bash
+python3 scripts/sanity_check.py | grep "warmstart/ours"
+# expect: warmstart/ours: 1 real / 55 stubs
+```