Delete generate_ik_parquets.py

generate_ik_parquets.py

@@ -1,135 +0,0 @@
-import os
-import pathlib
-import time
-import numpy as np
-import mujoco
-import pandas as pd
-from huggingface_hub import upload_file
-from so100_mujoco_sim import so100_ik, get_arm_joint_indices
-
-# --- Config ---
-N_EPISODES = 1
-STEPS_PER_PHASE = 60
-HF_REPO_ID = "arulloomba/so100_simulation_test"
-
-# --- Paths ---
-SCRIPT_DIR = pathlib.Path(__file__).parent
-MODEL_PATH = SCRIPT_DIR / "xml/sim_scene.xml"
-PARQUET_OUT_DIR = SCRIPT_DIR / "ik_parquets"
-PARQUET_OUT_DIR.mkdir(exist_ok=True)
-
-# --- Randomization bounds (meters) ---
-CUBE_X_RANGE = (0.00, 0.10)
-CUBE_Y_RANGE = (-0.28, -0.16)
-CUBE_Z = 0.0  # Table height for cube
-
-CUBE_BODY = "block_a"
-
-# --- Helper: set free joint pose ---
-def set_body_pos(model, data, body_name, pos):
-    body_id = mujoco.mj_name2id(model, mujoco.mjtObj.mjOBJ_BODY, body_name)
-    jnt_id = model.body_jntadr[body_id]
-    if model.jnt_type[jnt_id] != mujoco.mjtJoint.mjJNT_FREE:
-        raise ValueError(f"Body {body_name} does not have a free joint!")
-    qpos_adr = model.jnt_qposadr[jnt_id]
-    data.qpos[qpos_adr:qpos_adr+3] = pos
-    mujoco.mj_forward(model, data)
-
-# --- Main loop ---
-for ep in range(N_EPISODES):
-    model = mujoco.MjModel.from_xml_path(str(MODEL_PATH))
-    data = mujoco.MjData(model)
-    arm_joint_indices = get_arm_joint_indices(model)
-
-    # Randomize cube position
-    cube_x = np.random.uniform(*CUBE_X_RANGE)
-    cube_y = np.random.uniform(*CUBE_Y_RANGE)
-    set_body_pos(model, data, CUBE_BODY, [cube_x, cube_y, CUBE_Z])
-    mujoco.mj_forward(model, data)
-
-    # Compute adjusted cube center for IK
-    cube_center = np.array([cube_x, cube_y - 0.02, CUBE_Z])
-
-    # Reset arm to home
-    home_qpos = np.array([0, -3.14, 3.14, .87, 0, -0.20])
-    data.qpos[:6] = home_qpos
-    mujoco.mj_forward(model, data)
-
-    # --- PHASES: above, descend, grasp, lift, return ---
-    z_offset_above = 0.18
-    z_offset_grasp = 0.025
-    z_offset_lift = 0.18
-
-    # 1. Above cube
-    target_above = cube_center + np.array([0, 0, z_offset_above])
-    qpos_above = so100_ik(model, data, target_above, point_down=True)
-    qpos_above[3] = 1.8  # wrist pitch
-    qpos_above[5] = 0.7  # gripper open
-
-    # 2. Descend to grasp
-    target_grasp = cube_center + np.array([0, 0, z_offset_grasp])
-    qpos_grasp = so100_ik(model, data, target_grasp, point_down=True)
-    qpos_grasp[3] = 1.8
-    qpos_grasp[5] = 0.7  # gripper open
-
-    # 3. Close gripper
-    qpos_closed = qpos_grasp.copy()
-    qpos_closed[5] = -0.1  # gripper closed
-
-    # 4. Lift
-    target_lift = cube_center + np.array([0, 0, z_offset_lift])
-    qpos_lift = so100_ik(model, data, target_lift, point_down=True)
-    qpos_lift[3] = 1.8
-    qpos_lift[5] = -0.1  # gripper closed
-
-    # 5. Return to home
-    qpos_home = home_qpos.copy()
-
-    # --- Sequence ---
-    phase_qpos = [home_qpos, qpos_above, qpos_grasp, qpos_closed, qpos_lift, qpos_home]
-    phase_names = ["home", "above", "grasp", "closed", "lift", "return"]
-
-    records = []
-    sim_time = 0.0
-    dt = 0.01
-    frame_index = 0
-    for i in range(len(phase_qpos)-1):
-        start = phase_qpos[i]
-        end = phase_qpos[i+1]
-        for t in range(STEPS_PER_PHASE):
-            interp = (t + 1) / STEPS_PER_PHASE
-            target = start + (end - start) * interp
-            data.ctrl[:6] = target[:6]
-            mujoco.mj_step(model, data)
-            records.append({
-                "sequence.observation.state": data.qpos[:6].copy(),
-                "sequence.timestamp": float(sim_time),
-                "frame_index": frame_index,
-                "episode_index": ep,
-                "index": frame_index,
-                "task_index": 0,
-                "action": target[:6].copy(),
-                "cube_pos": [cube_x, cube_y, CUBE_Z],
-            })
-            frame_index += 1
-            sim_time += dt
-
-    # Save parquet
-    df = pd.DataFrame(records)
-    df["sequence.observation.state"] = df["sequence.observation.state"].apply(lambda x: x.tolist())
-    df["action"] = df["action"].apply(lambda x: x.tolist())
-    parquet_path = PARQUET_OUT_DIR / f"ik_ep_{ep:05d}.parquet"
-    df.to_parquet(parquet_path, index=False)
-    print(f"[Episode {ep+1}/{N_EPISODES}] Saved {parquet_path}")
-
-    # Upload to Hugging Face
-    upload_file(
-        repo_id=HF_REPO_ID,
-        path_or_fileobj=str(parquet_path),
-        path_in_repo=f"parquets/{parquet_path.name}",
-        repo_type="dataset"
-    )
-    print(f"[Episode {ep+1}/{N_EPISODES}] Uploaded to Hugging Face as parquets/{parquet_path.name}")
-    time.sleep(0.5)
-
-print("All episodes complete.")