Package exact adaptive implementation: code/examples/realRobots/Piper/eval_files/visualize_adaptive_bspline_open_loop.py

models/piper-pick-white-block-25hz-qwenpi-v3-adaptive-left-exclude-gripper-native-controls13-gb32-50k-20260905/code/examples/realRobots/Piper/eval_files/visualize_adaptive_bspline_open_loop.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+"""Visualize adaptive-vs-uniform open-loop errors on the common prefix."""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+
+ACTION_NAMES = ["J1", "J2", "J3", "J4", "J5", "J6", "Gripper"]
+
+
+def _rmse(values, axis):
+    return np.sqrt(np.mean(np.square(np.asarray(values, dtype=np.float64)), axis=axis))
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("adaptive_arrays", type=Path)
+    parser.add_argument("uniform_arrays", type=Path)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    with np.load(args.adaptive_arrays.expanduser().resolve()) as handle:
+        adaptive = {key: handle[key] for key in handle.files}
+    with np.load(args.uniform_arrays.expanduser().resolve()) as handle:
+        uniform = {key: handle[key] for key in handle.files}
+    np.testing.assert_array_equal(adaptive["sample_index"], uniform["sample_index"])
+    horizon = 10
+    uniform_model = uniform["prediction_vs_original_error"][:, :horizon]
+    uniform_representation = uniform["spline_vs_original_error"][:, :horizon]
+    adaptive_model = adaptive["prediction_vs_original_error"][:, :horizon]
+    adaptive_representation = adaptive["spline_vs_original_error"][:, :horizon]
+
+    fig, axes = plt.subplots(2, 2, figsize=(14, 9), constrained_layout=True)
+    steps = np.arange(horizon)
+    axes[0, 0].plot(steps, _rmse(adaptive_model[..., :6], axis=(0, 2)), marker="o", label="Adaptive")
+    axes[0, 0].plot(steps, _rmse(uniform_model[..., :6], axis=(0, 2)), marker="o", label="Uniform-left")
+    axes[0, 0].set(title="Policy arm RMSE by horizon", xlabel="25 Hz action step", ylabel="RMSE")
+    axes[0, 0].legend()
+    axes[0, 0].grid(alpha=0.25)
+
+    axes[0, 1].plot(steps, _rmse(adaptive_representation[..., :6], axis=(0, 2)), marker="o", label="Adaptive")
+    axes[0, 1].plot(steps, _rmse(uniform_representation[..., :6], axis=(0, 2)), marker="o", label="Uniform-left")
+    axes[0, 1].set(title="Representation-only arm RMSE", xlabel="25 Hz action step", ylabel="RMSE")
+    axes[0, 1].legend()
+    axes[0, 1].grid(alpha=0.25)
+
+    width = 0.36
+    dims = np.arange(7)
+    axes[1, 0].bar(dims - width / 2, _rmse(adaptive_model, axis=(0, 1)), width, label="Adaptive")
+    axes[1, 0].bar(dims + width / 2, _rmse(uniform_model, axis=(0, 1)), width, label="Uniform-left")
+    axes[1, 0].set_xticks(dims, ACTION_NAMES)
+    axes[1, 0].set(title="Policy RMSE per action dimension", ylabel="RMSE")
+    axes[1, 0].legend()
+    axes[1, 0].grid(axis="y", alpha=0.25)
+
+    duration_error = adaptive["duration_error"].reshape(-1)
+    bins = np.arange(duration_error.min() - 0.5, duration_error.max() + 1.5)
+    axes[1, 1].hist(duration_error, bins=bins, density=True, color="#5b8ff9", alpha=0.85)
+    axes[1, 1].set(
+        title="Adaptive predicted knot-duration error",
+        xlabel="Predicted − target duration (25 Hz steps)",
+        ylabel="Fraction",
+    )
+    axes[1, 1].grid(axis="y", alpha=0.25)
+    fig.suptitle(
+        f"Piper QwenPI_v3: adaptive-left vs uniform-left ({len(adaptive_model):,} observations)",
+        fontsize=15,
+    )
+
+    output = args.output.expanduser().resolve()
+    output.parent.mkdir(parents=True, exist_ok=True)
+    fig.savefig(output, dpi=180)
+    fig.savefig(output.with_suffix(".svg"))
+    plt.close(fig)
+    print(output)
+    print(output.with_suffix(".svg"))
+
+
+if __name__ == "__main__":
+    main()