"""Build fig4_ood.svg — OOD generalization, 2 rows × 3 cols.
Row 1: Spatial / object position generalization
Row 2: Viewpoint generalization
Each row: (a/d) train/test distribution, (b/e) chart vs PARA/ACT,
(c/f) qualitative PARA success vs ACT failure at an OOD condition.
Charts are generated with matplotlib (matching fig4b_pertheta style)
and embedded as PNGs into the SVG.
"""
import base64
import os
import time
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
_t = time.time()
PARA_GREEN = "#16653a"
ACT_RED = "#a12029"
TEXT_DARK = "#0f172a"
CHART_OUT = "/data/cameron/penpot/figures/extracted/fig4"
os.makedirs(CHART_OUT, exist_ok=True)
def _styled_chart_axes(ax):
"""Apply common fig4b-style axes styling."""
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)
def render_spatial_chart(path):
"""Row 1 middle chart: success rate vs distance from train boundary.
Reads /tmp/fig4_spatial_per_dx.json (per-position eval, 19 test positions
× 5 episodes each) and plots scatter + binned mean line for PARA and ACT.
"""
import json
with open("/tmp/fig4_spatial_per_dx.json") as f:
data = json.load(f)
results = data["results"] if isinstance(data, dict) and "results" in data else data
pts = sorted(results, key=lambda r: r["dx"])
dx_cm = [r["dx"] * 100 for r in pts] # meters → centimeters
para = [r["para"] * 100 for r in pts]
act = [r["act"] * 100 for r in pts]
fig, ax = plt.subplots(figsize=(6, 3.8), dpi=300)
for y in (25, 50, 75):
ax.axhline(y, color="#888888", alpha=0.3, linewidth=0.5, zorder=1)
# Boundary marker (train ↔ test divider at dx=0)
ax.axvline(0, color="#94a3b8", linestyle=":", linewidth=1.2, zorder=2)
# Scatter (raw per-position)
ax.scatter(dx_cm, para, s=42, color=PARA_GREEN, edgecolor="white",
linewidth=1.0, zorder=4, alpha=0.85)
ax.scatter(dx_cm, act, s=42, color=ACT_RED, edgecolor="white",
linewidth=1.0, marker="s", zorder=3, alpha=0.85)
# Binned mean line (5 bins across dx range)
import numpy as np
bins = np.linspace(min(dx_cm) - 0.1, max(dx_cm) + 0.1, 6)
bin_centers, bin_para, bin_act = [], [], []
for i in range(len(bins) - 1):
lo, hi = bins[i], bins[i + 1]
members_p = [p for d, p in zip(dx_cm, para) if lo <= d < hi]
members_a = [a for d, a in zip(dx_cm, act) if lo <= d < hi]
if members_p:
bin_centers.append((lo + hi) / 2)
bin_para.append(sum(members_p) / len(members_p))
bin_act.append(sum(members_a) / len(members_a))
ax.plot(bin_centers, bin_para, color=PARA_GREEN, linewidth=2.4,
marker="o", markersize=10, markerfacecolor=PARA_GREEN,
markeredgecolor="white", markeredgewidth=1.5,
label="PARA", zorder=6)
ax.plot(bin_centers, bin_act, color=ACT_RED, linewidth=2.4,
marker="s", markersize=10, markerfacecolor=ACT_RED,
markeredgecolor="white", markeredgewidth=1.5,
linestyle="--", label="ACT", zorder=5)
ax.set_xlabel("Distance from train boundary (cm)", fontsize=12)
ax.set_ylabel("Test-set Success Rate (%)", fontsize=12)
ax.set_yticks([0, 25, 50, 75, 100])
ax.set_ylim(-3, 105)
_styled_chart_axes(ax)
# "train side / test side" annotation
ax.text(min(dx_cm) - 1, 95, "train\nside", fontsize=9, color="#94a3b8",
ha="left", va="top", fontweight="700", linespacing=1.0)
ax.text(max(dx_cm) + 1, 95, "test\nside", fontsize=9, color="#94a3b8",
ha="right", va="top", fontweight="700", linespacing=1.0)
leg = ax.legend(loc="upper right", frameon=True, fontsize=10,
edgecolor="#cccccc", facecolor="white")
leg.get_frame().set_linewidth(0.8)
plt.tight_layout()
fig.savefig(path, format="png", bbox_inches="tight", dpi=300,
facecolor="white", edgecolor="none")
plt.close(fig)
def render_viewpoint_chart(path):
"""Row 2 middle chart: per-theta (same data as fig4b_pertheta)."""
fig, ax = plt.subplots(figsize=(6, 3.8), 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]
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=2.2, marker="o",
markersize=9, markerfacecolor=PARA_GREEN, markeredgecolor="white",
markeredgewidth=1.2, label="PARA", zorder=5)
ax.plot(theta, act, color=ACT_RED, linewidth=2.2, marker="s",
markersize=9, 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("Test-set Success Rate (%)", fontsize=12)
ax.set_xlim(-1, 26)
ax.set_ylim(0, 100)
ax.set_xticks(theta)
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])
_styled_chart_axes(ax)
leg = ax.legend(loc="upper right", frameon=True, fontsize=11,
edgecolor="#cccccc", facecolor="white")
leg.get_frame().set_linewidth(0.8)
plt.tight_layout()
fig.savefig(path, format="png", bbox_inches="tight", dpi=300,
facecolor="white", edgecolor="none")
plt.close(fig)
spatial_chart_path = f"{CHART_OUT}/spatial_chart.png"
viewpoint_chart_path = f"{CHART_OUT}/viewpoint_chart.png"
render_spatial_chart(spatial_chart_path)
render_viewpoint_chart(viewpoint_chart_path)
print(f"[{time.time()-_t:.2f}s] generated charts")
def b64(p):
return base64.b64encode(open(p, "rb").read()).decode()
DASHBOARD = "/data/cameron/para/.agents/reports/project_site/media"
# Images for the 6 cells
spatial_dist_b64 = b64(f"{DASHBOARD}/exp3_leftright_distribution.png")
vp_dist_b64 = b64(f"{CHART_OUT}/vp_polar_only.png")
spatial_chart_b64 = b64(spatial_chart_path)
viewpoint_chart_b64 = b64(viewpoint_chart_path)
act_spatial_b64 = b64("/data/cameron/penpot/figures/extracted/4c_act_pos_extreme.png")
para_spatial_b64 = b64("/data/cameron/penpot/figures/extracted/4c_para_pos_extreme.png")
act_vp_b64 = b64("/data/cameron/penpot/figures/extracted/4c_act_vp_extreme.png")
para_vp_b64 = b64("/data/cameron/penpot/figures/extracted/4c_para_vp_extreme.png")
GREEN = PARA_GREEN
RED = ACT_RED
GRAY = "#6b7280"
LIGHT = "#e5e7eb"
TXT = "#0f172a"
# ═══════════════════════════════════════════════════════════════════════════
# Layout: 1400 x 820, 2 rows × 3 cols
# Rows are 50px taller each so charts/frames have more vertical room.
# ═══════════════════════════════════════════════════════════════════════════
ROW1_Y = 50
ROW1_H = 360
ROW2_Y = 440
ROW2_H = 360
COL_X = [20, 480, 940]
COL_W = [440, 440, 440]
# Qualitative-frame geometry (side-by-side, square)
FRAME_SIZE = 200
FRAME_GAP = 10
def card(x, y, w, h, label, body, body_filter=True):
"""Wrap card: outer rect + sub-label + body SVG."""
return f'''
{label}
{body}
'''
# ── Row 1 cells ──
# (a) Spatial distribution — image fills most of the cell
r1c1_body = f'''
'''
# (b) Spatial chart
r1c2_body = f'''
'''
# (c) Spatial qualitative — two frames side-by-side with annotation circles
# Frame layout: ACT (left) and PARA (right), each 200x200, centered in the cell
_r1c3_x0 = COL_X[2] + (COL_W[2] - (2 * FRAME_SIZE + FRAME_GAP)) // 2
_r1c3_act_x = _r1c3_x0
_r1c3_para_x = _r1c3_x0 + FRAME_SIZE + FRAME_GAP
_r1c3_y = ROW1_Y + 50 # after card label
# Spatial frame approx circle positions (source 448x448, slice scale = 200/448 = 0.446)
# Visible image: entire source (square → square)
# ACT_fail: gripper is hovering ~center-upper area; target (plate) is center-lower
_sp_scale = FRAME_SIZE / 448
_sp_act_grip = (int(250 * _sp_scale), int(185 * _sp_scale)) # ACT gripper (wrong spot)
_sp_bowl = (int(325 * _sp_scale), int(330 * _sp_scale)) # target plate
_sp_para_grip = (int(250 * _sp_scale), int(270 * _sp_scale)) # PARA gripper (on target)
r1c3_body = f'''
ACT ✗
PARA ✓
ACT overshoots target · PARA reaches bowl
'''
# ── Row 2 cells ──
# (d) Viewpoint distribution
r2c1_body = f'''
'''
# (e) Viewpoint chart
r2c2_body = f'''
'''
# (f) Viewpoint qualitative — two frames side-by-side with annotation circles
# Viewpoint frames are ~448x373 (aspect 1.20). With slice into 200x200, scale = 0.536.
# Horizontal crop: (448*0.536 - 200)/2 = 20 px each side in SVG coords.
_r2c3_x0 = COL_X[2] + (COL_W[2] - (2 * FRAME_SIZE + FRAME_GAP)) // 2
_r2c3_act_x = _r2c3_x0
_r2c3_para_x = _r2c3_x0 + FRAME_SIZE + FRAME_GAP
_r2c3_y = ROW2_Y + 50
# Source pixel → SVG frame coord (slice mode, xMidYMid, 448x373 source)
_vp_scale = max(FRAME_SIZE / 448, FRAME_SIZE / 373) # slice = max scale
# Scaled image: 448*s x 373*s. Crop: horizontal (448*s - 200)/2, vertical (373*s - 200)/2
def _vp_svg(px, py):
sx = px * _vp_scale - (448 * _vp_scale - FRAME_SIZE) / 2
sy = py * _vp_scale - (373 * _vp_scale - FRAME_SIZE) / 2
return (sx, sy)
_vp_act_grip = _vp_svg(380, 110)
_vp_bowl = _vp_svg(385, 290)
_vp_para_grip = _vp_svg(385, 250)
r2c3_body = f'''
ACT ✗
PARA ✓
ACT collapses at new viewpoint · PARA tracks target
'''
svg = f'''
'''
out = "/data/cameron/para/paper/figs/svg/fig4_ood.svg"
with open(out, "w") as f:
f.write(svg)
print(f"[{time.time()-_t:.2f}s] wrote {out} ({len(svg)} bytes)")