Upload code/1027_run/tasks/lever2npz.py with huggingface_hub

code/1027_run/tasks/lever2npz.py

@@ -0,0 +1,73 @@
+#!/usr/bin/env python3
+"""Convert lever_sim.py output dir → ep_NNNNNN.npz in pack_zarr format.
+
+lever_sim writes:
+  <out>/ur5_steps.json  - list of {i,t,phase,q_mm,tip:[x,y,z],...}
+  <out>/ur5_cam0.mp4   - front cam (optional)
+  <out>/ur5_cam1.mp4   - wrist cam (optional)
+
+We produce:
+  <episodes_dir>/ep_NNNNNN.npz with keys:
+    robot0_eef_pos    (N,3) float32 -- tip xyz
+    robot0_eef_quat   (N,3) float32 -- rotvec (constant, derived from press orientation)
+    robot0_gripper_qpos (N,2) float32 -- [0.0, -0.0] (gripper always closed)
+    action            (N,7) float32 -- [eef_pos_next, eef_quat_next, gripper_next] = state[t+1]
+    timestamp         (N,)  float64
+
+The episode is accepted only if q_at_press_end_mm <= press_q_mm (lever actually pressed).
+"""
+import argparse, json, os, sys, glob
+import numpy as np
+
+ap = argparse.ArgumentParser()
+ap.add_argument('--src-dir', required=True, help='dir with ur5_steps.json from one lever_sim run')
+ap.add_argument('--episodes-dir', required=True, help='output episodes/ dir')
+ap.add_argument('--ep-idx', type=int, required=True, help='episode index for filename')
+ap.add_argument('--press-q-mm', type=float, default=-50.0, help='accept if q_at_press_end <= this (mm)')
+A = ap.parse_args()
+
+steps_file = os.path.join(A.src_dir, 'ur5_steps.json')
+result_file = os.path.join(A.src_dir, 'result.json')
+
+if not os.path.exists(steps_file):
+    sys.exit(f"No ur5_steps.json in {A.src_dir}")
+
+steps = json.load(open(steps_file))
+if not steps:
+    sys.exit("ur5_steps.json is empty")
+
+# Check result for success
+if os.path.exists(result_file):
+    res = json.load(open(result_file))
+    ur5 = res.get('ur5', {})
+    q_end = ur5.get('q_at_press_end_mm', None)
+    if q_end is not None and q_end > A.press_q_mm:
+        sys.exit(f"Episode rejected: q_at_press_end_mm={q_end:.1f} > {A.press_q_mm:.1f}")
+    print(f"q_at_press_end_mm={q_end}, accepted")
+
+tips = np.array([s['tip'] for s in steps], dtype=np.float32)  # (T, 3)
+T = len(tips)
+if T < 2:
+    sys.exit(f"Too few steps: {T}")
+
+# rotvec: constant (gripper always closed, orientation approximately fixed during press)
+# Use zero rotvec as placeholder — the DP will learn from tip trajectory
+rotvec = np.zeros((T, 3), dtype=np.float32)
+gripper = np.zeros((T, 2), dtype=np.float32)  # closed = 0 width
+
+# state: [pos(3), rotvec(3), gripper_w(1)] -> 7D
+state = np.concatenate([tips, rotvec, gripper[:, :1]], axis=1)  # (T,7)
+# action = state[t+1], state = state[t], both of length T-1
+N = T - 1
+ep = dict(
+    robot0_eef_pos=tips[:N],
+    robot0_eef_quat=rotvec[:N],
+    robot0_gripper_qpos=gripper[:N],
+    action=state[1:],  # next state
+    timestamp=(np.arange(N) * 0.1).astype(np.float64),
+)
+
+os.makedirs(A.episodes_dir, exist_ok=True)
+out = os.path.join(A.episodes_dir, f'ep_{A.ep_idx:06d}.npz')
+np.savez_compressed(out, **ep)
+print(f"Wrote {out} ({N} rows)")