"""Render all 6 paper figures as PNGs (and SVGs where pure-vector). Output: /data/cameron/para/paper/figs/generated/figN_*.png Run: python3 render_paper_figures.py [--figs 1,2,3,4,5,6 | all] """ import argparse import os import sys import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt import matplotlib.gridspec as gridspec import matplotlib.patches as patches import matplotlib.image as mpimg import numpy as np # ── Style ── plt.rcParams["font.family"] = "DejaVu Sans" plt.rcParams["font.size"] = 11 PARA_GREEN = "#16653a" ACT_RED = "#a12029" NEUTRAL = "#71717a" TEXT_DARK = "#1f2d3d" SUBTLE = "#5a5a5a" BG_LIGHT = "#f8f9fa" BORDER = "#d4d4d8" OUT_DIRS = [ "/data/cameron/para/paper/figs/generated", "/data/cameron/para/.agents/reports/paper_figures/media", ] EXTRACTED = "/data/cameron/penpot/figures/extracted" EXISTING_FIGS = "/data/cameron/para/paper/figs/figma" DASHBOARD_MEDIA = "/data/cameron/para/.agents/reports/project_site/media" for d in OUT_DIRS: os.makedirs(d, exist_ok=True) # ───────────────────────────────────────────────────────────────────────────── # Helpers # ───────────────────────────────────────────────────────────────────────────── def save(fig, name, also_svg=False): """Save to all output locations.""" for out_dir in OUT_DIRS: fig.savefig(f"{out_dir}/{name}.png", format="png", bbox_inches="tight", dpi=200, facecolor="white") if also_svg: fig.savefig(f"{out_dir}/{name}.svg", format="svg", bbox_inches="tight", facecolor="white") plt.close(fig) print(f" wrote {name}.png to {len(OUT_DIRS)} locations" + (" + .svg" if also_svg else "")) def axis_off(ax): ax.set_xticks([]) ax.set_yticks([]) for s in ax.spines.values(): s.set_visible(False) def rounded_box(ax, x, y, w, h, fill, edgecolor=None, linewidth=1.5, rad=0.06): box = patches.FancyBboxPatch( (x, y), w, h, boxstyle=f"round,pad=0,rounding_size={rad}", facecolor=fill, edgecolor=edgecolor or fill, linewidth=linewidth, transform=ax.transData, clip_on=False, zorder=2) ax.add_patch(box) return box def arrow(ax, x1, y1, x2, y2, color=TEXT_DARK, lw=1.6): ax.annotate("", xy=(x2, y2), xytext=(x1, y1), arrowprops=dict(arrowstyle="->", color=color, lw=lw), zorder=3) # ───────────────────────────────────────────────────────────────────────────── # Figure 1: Overview teaser # ───────────────────────────────────────────────────────────────────────────── def render_fig1(): """Fig 1: Overview teaser. Working draft — to be polished in Figma later. No top title (paper caption owns it). No italic punchline. """ fig = plt.figure(figsize=(12, 5.0), dpi=300) gs = gridspec.GridSpec(1, 3, figure=fig, left=0.04, right=0.96, top=0.96, bottom=0.04, wspace=0.18, width_ratios=[1.4, 1.0, 1.0]) # ── Panel (a): Architecture comparison ── axA = fig.add_subplot(gs[0, 0]) axA.set_xlim(0, 10) axA.set_ylim(0, 6.5) axis_off(axA) axA.text(5, 6.2, "(a) Architecture Comparison", ha="center", va="bottom", fontsize=12, fontweight="700") # Global Regression row axA.text(0.2, 5.4, "Global Regression (ACT)", ha="left", va="center", fontsize=9.5, fontweight="700", color=ACT_RED) g_boxes = [("Image", "#eaf2fb"), ("DINO", "#fff3e6"), ("CLS", "#f5e6fb"), ("MLP", "#ffffff"), ("(x,y,z)", "#fde9ec")] g_x = 0.15 g_y = 4.3 bw, bh = 1.45, 0.7 gap = 0.4 for i, (label, fill) in enumerate(g_boxes): bx = g_x + i * (bw + gap) rounded_box(axA, bx, g_y, bw, bh, fill, edgecolor=NEUTRAL) axA.text(bx + bw/2, g_y + bh/2, label, ha="center", va="center", fontsize=9, fontweight="600", zorder=3) if i < len(g_boxes) - 1: arrow(axA, bx + bw + 0.05, g_y + bh/2, bx + bw + gap - 0.05, g_y + bh/2) # PARA row axA.text(0.2, 3.0, "PARA (ours)", ha="left", va="center", fontsize=9.5, fontweight="700", color=PARA_GREEN) p_boxes = [("Image", "#eaf2fb"), ("DINO", "#fff3e6"), ("Heatmap\nVolume", "#fde9ec"), ("Argmax", "#ffffff"), ("3D Point", "#e8f5ec")] p_y = 1.9 for i, (label, fill) in enumerate(p_boxes): bx = g_x + i * (bw + gap) rounded_box(axA, bx, p_y, bw, bh, fill, edgecolor=NEUTRAL) axA.text(bx + bw/2, p_y + bh/2, label, ha="center", va="center", fontsize=9, fontweight="600", zorder=3) if i < len(p_boxes) - 1: arrow(axA, bx + bw + 0.05, p_y + bh/2, bx + bw + gap - 0.05, p_y + bh/2) # (italic punchline removed — paper caption owns it) # ── Panel (b): Vignettes ── axB = fig.add_subplot(gs[0, 1]) axB.set_xlim(0, 10) axB.set_ylim(0, 6.5) axis_off(axB) axB.text(5, 6.2, "(b) Where PARA Helps", ha="center", va="bottom", fontsize=12, fontweight="700") vignettes = [ ("OOD Generalization", "Shifted camera & objects", PARA_GREEN), ("Video Backbone", "SVD video → robot", "#3b6fa6"), ("Cross-Embodiment", "Arm-deleted point tracks", "#5a3da2"), ] vy_start = 5.2 vh = 1.5 for i, (title, sub, color) in enumerate(vignettes): vy = vy_start - i * vh # Icon circle circle = patches.Circle((0.9, vy), 0.4, facecolor=color, edgecolor="white", linewidth=2, zorder=3) axB.add_patch(circle) axB.text(1.7, vy + 0.15, title, ha="left", va="center", fontsize=11, fontweight="700") axB.text(1.7, vy - 0.25, sub, ha="left", va="center", fontsize=9, color=SUBTLE) # ── Panel (c): Headlines ── axC = fig.add_subplot(gs[0, 2]) axC.set_xlim(0, 10) axC.set_ylim(0, 6.5) axis_off(axC) axC.text(5, 6.2, "(c) Headline Results", ha="center", va="bottom", fontsize=12, fontweight="700") headlines = [ ("97% vs 9%", "Real Robot (20 demos)", PARA_GREEN), ("90% vs 0%", "Video Backbone", PARA_GREEN), ("[TBD]", "Point Track Pretraining", NEUTRAL), ] hy_start = 5.0 hh = 1.55 for i, (big, sub, color) in enumerate(headlines): hy = hy_start - i * hh # Card background rounded_box(axC, 0.5, hy - 0.55, 9.0, 1.25, "#ffffff", edgecolor=BORDER, linewidth=1.2, rad=0.1) axC.text(5, hy + 0.15, big, ha="center", va="center", fontsize=22, fontweight="800", color=color) axC.text(5, hy - 0.4, sub, ha="center", va="center", fontsize=10, color=SUBTLE, fontweight="500") save(fig, "fig1_overview") # ───────────────────────────────────────────────────────────────────────────── # Figure 2: Method details # ───────────────────────────────────────────────────────────────────────────── def render_fig2(): """Fig 2: Method pipeline + height illustration. No top title (paper caption owns it).""" fig = plt.figure(figsize=(12, 5.6), dpi=300) # No suptitle — paper caption handles it gs = gridspec.GridSpec(1, 2, figure=fig, left=0.02, right=0.98, top=0.92, bottom=0.04, wspace=0.06, width_ratios=[1.45, 1.0]) # b panel ~40% width for readable text # Sub-titles (smaller, lighter) fig.text(0.30, 0.96, "(a) Inference pipeline visualization", fontsize=11, fontweight="600", color="#555555", ha="center") fig.text(0.76, 0.96, "(b) Height vs Depth", fontsize=11, fontweight="600", color="#555555", ha="center") # ── Panel (a): 4-frame method strip with arrows between ── # Use a 7-column grid: img | arrow | img | arrow | img | arrow | img gsA = gridspec.GridSpecFromSubplotSpec( 1, 7, subplot_spec=gs[0, 0], width_ratios=[4, 0.5, 4, 0.5, 4, 0.5, 4], wspace=0.0) captions = ["Camera Frustum", "Heatmap Volume", "Argmax → 3D Target", "Robot at Target"] paths = [ f"{EXTRACTED}/frame_2_stage1_clean.png", f"{EXTRACTED}/frame_2_stage2_clean.png", f"{EXTRACTED}/frame_2_stage3_clean.png", f"{EXTRACTED}/frame_2_stage4_clean.png", ] img_cols = [0, 2, 4, 6] arrow_cols = [1, 3, 5] for i, (col, p, cap) in enumerate(zip(img_cols, paths, captions)): ax = fig.add_subplot(gsA[0, col]) img = mpimg.imread(p) ax.imshow(img) ax.set_xticks([]) ax.set_yticks([]) # Light gray border around each frame for s in ax.spines.values(): s.set_visible(True) s.set_color("#cccccc") s.set_linewidth(1.0) ax.set_title(cap, fontsize=10, fontweight="600", color=SUBTLE, pad=4) for col in arrow_cols: ax = fig.add_subplot(gsA[0, col]) ax.set_xlim(0, 1) ax.set_ylim(0, 1) axis_off(ax) ax.annotate("", xy=(0.92, 0.5), xytext=(0.08, 0.5), arrowprops=dict( arrowstyle="-|>,head_length=0.7,head_width=0.5", color="#1f2d3d", lw=2.8)) # ── Panel (b): Height illustration ── axB = fig.add_subplot(gs[0, 1]) img_h = mpimg.imread(f"{EXISTING_FIGS}/height_illustration.png") axB.imshow(img_h) axB.set_xticks([]) axB.set_yticks([]) for s in axB.spines.values(): s.set_visible(True) s.set_color("#cccccc") s.set_linewidth(1.0) save(fig, "fig2_method") # ───────────────────────────────────────────────────────────────────────────── # Figure 3: Real robot results — single image # ───────────────────────────────────────────────────────────────────────────── def render_fig3(): img = mpimg.imread(f"{EXISTING_FIGS}/para_results.png") h, w = img.shape[:2] aspect = h / w fig_w = 12 fig_h = fig_w * aspect + 0.6 # extra space for title fig = plt.figure(figsize=(fig_w, fig_h), dpi=200) fig.suptitle("Figure 3: Real Robot Results (SO-100, 20 demos)", fontsize=16, fontweight="700", y=0.99, color=TEXT_DARK) ax = fig.add_subplot(111) ax.imshow(img) axis_off(ax) plt.tight_layout(rect=[0, 0, 1, 0.96]) save(fig, "fig3_realrobot") # ───────────────────────────────────────────────────────────────────────────── # Figure 4: OOD Analysis (3 panels) # ───────────────────────────────────────────────────────────────────────────── def render_fig4(): """Composite Fig 4: stacked (a)+(b) on left, (c) on right (taller layout).""" fig = plt.figure(figsize=(12, 7.5), dpi=300) # No suptitle — paper caption owns it gs = gridspec.GridSpec(1, 2, figure=fig, left=0.04, right=0.985, top=0.96, bottom=0.05, wspace=0.13, width_ratios=[1.05, 1.0]) # Left column: stack (a) on top, (b) on bottom gsL = gridspec.GridSpecFromSubplotSpec(2, 1, subplot_spec=gs[0, 0], hspace=0.40, height_ratios=[1.0, 1.05]) # Sub-labels — (a) nudged left to avoid overlap with distribution plot's # built-in legend text ("Left half (j=0~7)...") fig.text(0.07, 0.965, "(a) Spatial Extrapolation", fontsize=12, fontweight="600", color="#555555", ha="left") fig.text(0.07, 0.488, "(b) Per-Angle Viewpoint Robustness", fontsize=12, fontweight="600", color="#555555", ha="left") fig.text(0.555, 0.965, "(c) Qualitative Comparison", fontsize=12, fontweight="600", color="#555555", ha="left") # ── Panel (a): distribution image + bar chart side-by-side inside ── gsA = gridspec.GridSpecFromSubplotSpec(1, 2, subplot_spec=gsL[0, 0], wspace=0.25, width_ratios=[1.4, 1.0]) axA1 = fig.add_subplot(gsA[0, 0]) img = mpimg.imread(f"{DASHBOARD_MEDIA}/exp3_leftright_distribution.png") axA1.imshow(img) axis_off(axA1) axA2 = fig.add_subplot(gsA[0, 1]) for y in (25, 50): axA2.axhline(y, color="#888888", alpha=0.3, linewidth=0.5, zorder=0) bars = axA2.bar(["PARA", "ACT"], [54, 1], color=[PARA_GREEN, ACT_RED], width=0.55) for b, val in zip(bars, [54, 1]): axA2.annotate(f"{val}%", xy=(b.get_x() + b.get_width()/2, b.get_height()), xytext=(0, 3), textcoords="offset points", ha="center", va="bottom", fontsize=12, fontweight="700") axA2.set_ylim(0, 70) axA2.set_ylabel("Success Rate (%)", fontsize=10) axA2.tick_params(labelsize=10, colors="#444444") axA2.spines["top"].set_visible(False) axA2.spines["right"].set_visible(False) axA2.spines["left"].set_color("#888888") axA2.spines["bottom"].set_color("#888888") # ── Panel (b): per-theta chart ── axB = fig.add_subplot(gsL[1, 0]) theta = [0, 3.6, 7.1, 10.7, 14.3, 17.9, 21.4, 25] para = [88, 79, 62, 63, 62, 62, 33, 38] act = [67, 54, 42, 17, 12, 0, 0, 0] for y in (25, 50, 75): axB.axhline(y, color="#888888", alpha=0.3, linewidth=0.5, zorder=1) axB.plot(theta, para, color=PARA_GREEN, linewidth=2, marker="o", markersize=8, markerfacecolor=PARA_GREEN, markeredgecolor="white", markeredgewidth=1.2, label="PARA", zorder=5) axB.plot(theta, act, color=ACT_RED, linewidth=2, marker="s", markersize=8, markerfacecolor=ACT_RED, markeredgecolor="white", markeredgewidth=1.2, linestyle="--", label="ACT", zorder=4) axB.set_xlabel("Camera Elevation θ (degrees)", fontsize=11) axB.set_ylabel("Success Rate (%)", fontsize=11) axB.set_xlim(-1, 26) axB.set_ylim(0, 100) axB.set_xticks(theta) axB.set_xticklabels(["0°\n(train)", "3.6", "7.1", "10.7", "14.3", "17.9", "21.4", "25"]) axB.set_yticks([0, 25, 50, 75, 100]) axB.spines["top"].set_visible(False) axB.spines["right"].set_visible(False) axB.spines["left"].set_color("#888888") axB.spines["bottom"].set_color("#888888") axB.tick_params(labelsize=10, colors="#444444") leg = axB.legend(loc="upper right", frameon=True, fontsize=10, edgecolor="#cccccc", facecolor="white") leg.get_frame().set_linewidth(0.8) # ── Panel (c): 2x2 qualitative grid with annotations ── gsC = gridspec.GridSpecFromSubplotSpec(2, 2, subplot_spec=gs[0, 1], hspace=0.14, wspace=0.04) ACT_GRIPPER_MID = (380, 110) BOWL_POS = (385, 290) PARA_GRIPPER_MID = (385, 230) cells = [ ("4c_act_start_v2.png", "ACT — start", ACT_RED, None), ("4c_act_mid_v2.png", "ACT — mid", ACT_RED, "act"), ("4c_para_start_v2.png", "PARA — start", PARA_GREEN, None), ("4c_para_mid_v2.png", "PARA — mid", PARA_GREEN, "para"), ] for i, (fn, label, color, annot) in enumerate(cells): ax = fig.add_subplot(gsC[i // 2, i % 2]) img = mpimg.imread(f"{EXTRACTED}/{fn}") ax.imshow(img) ax.set_xticks([]) ax.set_yticks([]) for s in ax.spines.values(): s.set_visible(True) s.set_color("#cccccc") s.set_linewidth(1.0) ax.set_title(label, fontsize=10, fontweight="700", color=color, pad=3) if annot == "act": ax.add_patch(patches.Circle(ACT_GRIPPER_MID, 36, facecolor="none", edgecolor=ACT_RED, linewidth=2.5, zorder=5)) ax.add_patch(patches.Circle(BOWL_POS, 36, facecolor="none", edgecolor=PARA_GREEN, linewidth=2.5, zorder=5)) elif annot == "para": ax.add_patch(patches.Circle(PARA_GRIPPER_MID, 42, facecolor="none", edgecolor=PARA_GREEN, linewidth=2.5, zorder=5)) save(fig, "fig4_ood") # ───────────────────────────────────────────────────────────────────────────── # Figure 5: Video Backbone # ───────────────────────────────────────────────────────────────────────────── def render_fig5(): """Fig 5: Two-stage training diagram + rollout grid comparison.""" fig = plt.figure(figsize=(12, 5.0), dpi=300) # No suptitle — paper caption owns it gs = gridspec.GridSpec(1, 2, figure=fig, left=0.03, right=0.97, top=0.90, bottom=0.05, wspace=0.08, width_ratios=[1.0, 1.4]) # Sub-titles (smaller, lighter — like Fig 2) fig.text(0.23, 0.95, "(a) Two-Stage Training", fontsize=11, fontweight="600", color="#555555", ha="center") fig.text(0.71, 0.95, "(b) Rollout Comparison", fontsize=11, fontweight="600", color="#555555", ha="center") # ── Panel (a): two-stage diagram ── axA = fig.add_subplot(gs[0, 0]) axA.set_xlim(0, 10) axA.set_ylim(0, 6) axis_off(axA) # Box 1: Video UNet rounded_box(axA, 0.8, 2.5, 3.5, 1.6, "#eaf2fb", edgecolor="#3b6fa6", linewidth=2, rad=0.15) axA.text(2.55, 3.6, "Video UNet", ha="center", va="center", fontsize=12, fontweight="700") axA.text(2.55, 3.0, "(4K pretrain)", ha="center", va="center", fontsize=10, color=SUBTLE) # Arrow axA.annotate("", xy=(5.4, 3.3), xytext=(4.4, 3.3), arrowprops=dict( arrowstyle="-|>,head_length=0.6,head_width=0.4", color="#1f2d3d", lw=2.5)) # Box 2: Joint Fine-tune rounded_box(axA, 5.5, 2.5, 3.7, 1.6, "#fff3e6", edgecolor="#d97e1f", linewidth=2, rad=0.15) axA.text(7.35, 3.6, "Joint Fine-tune", ha="center", va="center", fontsize=12, fontweight="700") axA.text(7.35, 3.0, "(3K)", ha="center", va="center", fontsize=10, color=SUBTLE) # Two output branches — labels positioned away from box axA.annotate("PARA Heatmap Head", xy=(9.6, 4.7), xytext=(9.6, 5.4), fontsize=10, fontweight="600", color=PARA_GREEN, ha="center", arrowprops=dict( arrowstyle="-|>,head_length=0.5,head_width=0.3", color=PARA_GREEN, lw=1.6)) axA.annotate("Video Generation", xy=(9.6, 1.9), xytext=(9.6, 1.2), fontsize=10, fontweight="600", color=SUBTLE, ha="center", arrowprops=dict( arrowstyle="-|>,head_length=0.5,head_width=0.3", color=SUBTLE, lw=1.6)) # ── Panel (b): rollout comparison ── gsB = gridspec.GridSpecFromSubplotSpec(1, 2, subplot_spec=gs[0, 1], wspace=0.06) para_img = mpimg.imread(f"{EXTRACTED}/frame_5_para.png") glob_img = mpimg.imread(f"{EXTRACTED}/frame_5_global.png") axB1 = fig.add_subplot(gsB[0, 0]) axB1.imshow(para_img) axB1.set_xticks([]) axB1.set_yticks([]) for s in axB1.spines.values(): s.set_visible(True) s.set_color("#cccccc") s.set_linewidth(1.0) axB1.set_title("SVD + PARA — 92%", fontsize=12, fontweight="700", color=PARA_GREEN, pad=4) axB2 = fig.add_subplot(gsB[0, 1]) axB2.imshow(glob_img) axB2.set_xticks([]) axB2.set_yticks([]) for s in axB2.spines.values(): s.set_visible(True) s.set_color("#cccccc") s.set_linewidth(1.0) axB2.set_title("SVD + Global Regression — 0%", fontsize=12, fontweight="700", color=ACT_RED, pad=4) save(fig, "fig5_video") # ───────────────────────────────────────────────────────────────────────────── # Figure 6: Point Track Pretraining (placeholder) # ───────────────────────────────────────────────────────────────────────────── def render_fig6(): fig = plt.figure(figsize=(12, 4.5), dpi=200) fig.suptitle("Figure 6: Point Track Pretraining (Preliminary)", fontsize=16, fontweight="700", y=0.99, color=TEXT_DARK) gs = gridspec.GridSpec(1, 2, figure=fig, left=0.04, right=0.96, top=0.86, bottom=0.10, wspace=0.18, width_ratios=[1.2, 1.0]) # ── Panel (a): 3 placeholder frames ── gsA = gridspec.GridSpecFromSubplotSpec(1, 3, subplot_spec=gs[0, 0], wspace=0.08) for i in range(3): ax = fig.add_subplot(gsA[0, i]) ax.set_xlim(0, 10) ax.set_ylim(0, 7) axis_off(ax) # Tan table table = patches.Rectangle((1, 1), 8, 5, facecolor="#e8d5b7", edgecolor=BORDER, linewidth=1) ax.add_patch(table) # Object (gray) obj = patches.Ellipse((4.0, 3.5), 1.2, 0.6, facecolor="#888888") ax.add_patch(obj) # Green dot (EEF position) — moves across frames dot_x = 5.0 + i * 0.7 dot_y = 3.8 - i * 0.3 dot = patches.Circle((dot_x, dot_y), 0.35, facecolor=PARA_GREEN, edgecolor="white", linewidth=1.5) ax.add_patch(dot) ax.set_title(f"t = {i*5}", fontsize=10, color=SUBTLE, pad=2) fig.text(0.295, 0.92, "(a) Arm-Deleted Training Data", fontsize=12, fontweight="700", ha="center") fig.text(0.295, 0.04, "(robot arm invisible — only EEF position retained as supervision)", fontsize=9, color=SUBTLE, ha="center", style="italic") # ── Panel (b): placeholder bar chart ── axB = fig.add_subplot(gs[0, 1]) labels = ["PARA\npretrain", "PARA\nscratch", "Global\npretrain", "Global\nscratch"] values = [70, 30, 20, 15] colors = [PARA_GREEN, PARA_GREEN, ACT_RED, ACT_RED] alphas = [1.0, 0.4, 1.0, 0.4] bars = axB.bar(range(len(labels)), values, color=colors, width=0.65, edgecolor="white", linewidth=1.2) for b, a in zip(bars, alphas): b.set_alpha(a) for b, v in zip(bars, values): axB.annotate(f"{v}%", xy=(b.get_x() + b.get_width()/2, b.get_height()), xytext=(0, 3), textcoords="offset points", ha="center", va="bottom", fontsize=11, fontweight="700") axB.set_xticks(range(len(labels))) axB.set_xticklabels(labels, fontsize=9.5) axB.set_ylabel("Success Rate (%)", fontsize=10) axB.set_ylim(0, 100) axB.set_yticks([0, 25, 50, 75, 100]) axB.tick_params(labelsize=9, colors="#444444") axB.spines["top"].set_visible(False) axB.spines["right"].set_visible(False) axB.spines["left"].set_color("#888888") axB.spines["bottom"].set_color("#888888") axB.set_title("(b) Pretrain → Fine-tune (preliminary)", fontsize=12, fontweight="700", pad=6) fig.text(0.74, 0.04, "Placeholder values — final results pending from backbones agent", fontsize=9, color=NEUTRAL, ha="center", fontweight="500") save(fig, "fig6_pointtrack") # ───────────────────────────────────────────────────────────────────────────── # Figure 4b standalone (already exists; re-render for consistency) # ───────────────────────────────────────────────────────────────────────────── def render_fig4a_spatial(): """Standalone Fig 4a: distribution plot + bar chart, side-by-side.""" fig = plt.figure(figsize=(8, 4), dpi=300) gs = gridspec.GridSpec(1, 2, figure=fig, width_ratios=[1.4, 1.0], wspace=0.30, top=0.86, bottom=0.14, left=0.06, right=0.96) ax1 = fig.add_subplot(gs[0, 0]) img = mpimg.imread(f"{DASHBOARD_MEDIA}/exp3_leftright_distribution.png") ax1.imshow(img) axis_off(ax1) ax2 = fig.add_subplot(gs[0, 1]) bars = ax2.bar(["PARA", "ACT"], [54, 1], color=[PARA_GREEN, ACT_RED], width=0.55) for b, val in zip(bars, [54, 1]): ax2.annotate(f"{val}%", xy=(b.get_x() + b.get_width()/2, b.get_height()), xytext=(0, 4), textcoords="offset points", ha="center", va="bottom", fontsize=13, fontweight="700") ax2.set_ylim(0, 70) ax2.set_ylabel("Success Rate (%)", fontsize=11) ax2.tick_params(labelsize=11, colors="#444444") ax2.spines["top"].set_visible(False) ax2.spines["right"].set_visible(False) ax2.spines["left"].set_color("#888888") ax2.spines["bottom"].set_color("#888888") # Subtle gridlines like 4b for y in (25, 50): ax2.axhline(y, color="#888888", alpha=0.3, linewidth=0.5, zorder=0) fig.suptitle("Fig 4a: Spatial Extrapolation (train left, test right)", fontsize=13, fontweight="700", y=0.98, color=TEXT_DARK) for out_dir in OUT_DIRS: fig.savefig(f"{out_dir}/fig4a_spatial.png", format="png", bbox_inches="tight", dpi=300, facecolor="white", edgecolor="none") plt.close(fig) print(" wrote fig4a_spatial.png to 2 locations (dpi=300)") def render_fig4b_standalone(): fig, ax = plt.subplots(figsize=(6, 3.5), dpi=300) theta = [0, 3.6, 7.1, 10.7, 14.3, 17.9, 21.4, 25] para = [88, 79, 62, 63, 62, 62, 33, 38] act = [67, 54, 42, 17, 12, 0, 0, 0] # Subtle horizontal gridlines at 25/50/75 (zorder under data) for y in (25, 50, 75): ax.axhline(y, color="#888888", alpha=0.3, linewidth=0.5, zorder=1) ax.plot(theta, para, color=PARA_GREEN, linewidth=2, marker="o", markersize=8, markerfacecolor=PARA_GREEN, markeredgecolor="white", markeredgewidth=1.2, label="PARA", zorder=5) ax.plot(theta, act, color=ACT_RED, linewidth=2, marker="s", markersize=8, markerfacecolor=ACT_RED, markeredgecolor="white", markeredgewidth=1.2, linestyle="--", label="ACT", zorder=4) ax.set_xlabel("Camera Elevation Angle θ (degrees)", fontsize=12) ax.set_ylabel("Success Rate (%)", fontsize=12) ax.set_xlim(-1, 26) ax.set_ylim(0, 100) ax.set_xticks(theta) # Custom x-tick labels with "(train)" annotation on the 0° tick ax.set_xticklabels(["0°\n(train)", "3.6", "7.1", "10.7", "14.3", "17.9", "21.4", "25"]) ax.set_yticks([0, 25, 50, 75, 100]) ax.spines["top"].set_visible(False) ax.spines["right"].set_visible(False) ax.spines["left"].set_color("#888888") ax.spines["bottom"].set_color("#888888") ax.tick_params(colors="#444444", labelsize=11) leg = ax.legend(loc="upper right", frameon=True, fontsize=10, edgecolor="#cccccc", facecolor="white") leg.get_frame().set_linewidth(0.8) plt.tight_layout() # Save with white background (no frame), high dpi for out_dir in OUT_DIRS: fig.savefig(f"{out_dir}/fig4b_pertheta.png", format="png", bbox_inches="tight", dpi=300, facecolor="white", edgecolor="none") fig.savefig(f"{out_dir}/fig4b_pertheta.svg", format="svg", bbox_inches="tight", facecolor="white", edgecolor="none") plt.close(fig) print(" wrote fig4b_pertheta.png + .svg to 2 locations (dpi=300)") def render_fig4c_qualitative(): """Fig 4c: 2x2 qualitative comparison with annotations highlighting divergence. Frames at t=0 (start) and t=12 (mid-rollout, clearest divergence). Annotations: - ACT-mid: red circle on its gripper (wrong location), green circle on bowl - PARA-mid: green circle on its gripper (correct, at bowl) """ fig = plt.figure(figsize=(7.5, 5.5), dpi=300) gs = gridspec.GridSpec(2, 2, figure=fig, hspace=0.14, wspace=0.04, top=0.95, bottom=0.04, left=0.04, right=0.96) # Approximate gripper / bowl positions in the cropped frame coordinate space # (frames are ~448 wide, ~370 tall after cropping) # ACT mid: gripper hovers upper-right, NOT touching bowl # PARA mid: gripper extends down to bowl # Bowl/plate: bottom-right area of all frames ACT_GRIPPER_MID = (380, 110) # where ACT gripper actually is (upper right) BOWL_POS = (385, 290) # where the bowl actually is (lower right) PARA_GRIPPER_MID = (385, 230) # where PARA gripper is (close to bowl) cells = [ ("4c_act_start_v2.png", "ACT — start (t=0)", ACT_RED, None, None), ("4c_act_mid_v2.png", "ACT — mid (t=12)", ACT_RED, "act", None), ("4c_para_start_v2.png", "PARA — start (t=0)", PARA_GREEN, None, None), ("4c_para_mid_v2.png", "PARA — mid (t=12)", PARA_GREEN, None, "para"), ] for i, (fn, label, color, annot_act, annot_para) in enumerate(cells): ax = fig.add_subplot(gs[i // 2, i % 2]) img = mpimg.imread(f"{EXTRACTED}/{fn}") ax.imshow(img) ax.set_xticks([]) ax.set_yticks([]) for s in ax.spines.values(): s.set_visible(True) s.set_color("#cccccc") s.set_linewidth(1.0) ax.set_title(label, fontsize=12, fontweight="700", color=color, pad=4) # ACT-mid annotations: red on wrong gripper, green on correct bowl if annot_act: # Red circle on ACT gripper (wrong location) c1 = patches.Circle(ACT_GRIPPER_MID, 36, facecolor="none", edgecolor=ACT_RED, linewidth=3, zorder=5) ax.add_patch(c1) ax.annotate("ACT gripper\n(wrong)", xy=ACT_GRIPPER_MID, xytext=(180, 70), fontsize=9, fontweight="600", color=ACT_RED, ha="center", arrowprops=dict(arrowstyle="-", color=ACT_RED, lw=1.2)) # Green circle on bowl (correct location) c2 = patches.Circle(BOWL_POS, 36, facecolor="none", edgecolor=PARA_GREEN, linewidth=3, zorder=5) ax.add_patch(c2) ax.annotate("bowl\n(target)", xy=BOWL_POS, xytext=(180, 290), fontsize=9, fontweight="600", color=PARA_GREEN, ha="center", arrowprops=dict(arrowstyle="-", color=PARA_GREEN, lw=1.2)) # PARA-mid annotation: green circle around gripper which IS at bowl if annot_para: c = patches.Circle(PARA_GRIPPER_MID, 42, facecolor="none", edgecolor=PARA_GREEN, linewidth=3, zorder=5) ax.add_patch(c) ax.annotate("PARA gripper\nreaches bowl", xy=PARA_GRIPPER_MID, xytext=(180, 200), fontsize=9, fontweight="600", color=PARA_GREEN, ha="center", arrowprops=dict(arrowstyle="-", color=PARA_GREEN, lw=1.2)) # No suptitle — paper caption owns it for out_dir in OUT_DIRS: fig.savefig(f"{out_dir}/fig4c_qualitative.png", format="png", bbox_inches="tight", dpi=300, facecolor="white", edgecolor="none") plt.close(fig) print(" wrote fig4c_qualitative.png to 2 locations (dpi=300)") BUILDERS = { "1": render_fig1, "2": render_fig2, "3": render_fig3, "4": render_fig4, "4a": render_fig4a_spatial, "4b": render_fig4b_standalone, "4c": render_fig4c_qualitative, "5": render_fig5, "6": render_fig6, } def main(): parser = argparse.ArgumentParser() parser.add_argument("--figs", default="all") args = parser.parse_args() if args.figs == "all": keys = ["1", "2", "3", "4", "4b", "5", "6"] else: keys = [k.strip() for k in args.figs.split(",")] for k in keys: if k not in BUILDERS: print(f"unknown: {k}") sys.exit(1) print(f"=== render fig {k} ===") BUILDERS[k]() if __name__ == "__main__": main()