"""Build paper figures in the para_paper_figures Penpot file. Each figure = one frame on the page. Uses a hybrid of: - Native rect/text shapes (editable in Penpot) - Image media uploads + image shape placement (for charts, photos, video frames) Run with --figs to specify which figures to build: python3 build_paper_figures.py --figs 4b python3 build_paper_figures.py --figs all python3 build_paper_figures.py --figs 1,2,3 """ import argparse import json import subprocess import sys import uuid from pathlib import Path # ── Penpot identifiers ── FILE_ID = "5688a1b5-92ba-8001-8007-dc49a465085a" PAGE_ID = "5688a1b5-92ba-8001-8007-dc49a465085b" ROOT = "00000000-0000-0000-0000-000000000000" COOKIE = "/tmp/penpot_user_cookies.txt" BASE_URL = "https://penpot.omidlab.net" # ── Color palette ── PARA_GREEN = "#16653a" ACT_RED = "#a12029" NEUTRAL = "#71717a" TEXT_DARK = "#1f2d3d" SUBTLE = "#5a5a5a" BG_LIGHT = "#f8f9fa" BORDER = "#d4d4d8" # ── Frame layout ── FRAME_W = 1200 FRAME_H_DEFAULT = 600 FRAME_GAP = 80 # vertical gap between figure frames FRAME_X = 100 # left margin # Cache the current revn so we can chain calls _revn_cache = None # ───────────────────────────────────────────────────────────────────────────── # Penpot API helpers # ───────────────────────────────────────────────────────────────────────────── def get_revn(): r = subprocess.run([ "curl", "-s", "-b", COOKIE, f"{BASE_URL}/api/rpc/command/get-file?id={FILE_ID}", "-H", "Accept: application/json", ], capture_output=True, text=True) return json.loads(r.stdout)["revn"] def upload_media(file_path, name): """Upload an image to the Penpot file. Returns (media_id, width, height, mtype).""" print(f" Uploading {name} ({Path(file_path).name})...") r = subprocess.run([ "curl", "-s", "-b", COOKIE, "-X", "POST", f"{BASE_URL}/api/rpc/command/upload-file-media-object", "-H", "Accept: application/json", "-F", f"file-id={FILE_ID}", "-F", "is-local=true", "-F", f"name={name}", "-F", f"content=@{file_path};type={_mtype_for(file_path)}", ], capture_output=True, text=True) d = json.loads(r.stdout) if "id" not in d: raise RuntimeError(f"upload failed: {r.stdout[:300]}") return d["id"], d["width"], d["height"], d["mtype"] def _mtype_for(path): p = str(path).lower() if p.endswith(".png"): return "image/png" if p.endswith(".jpg") or p.endswith(".jpeg"): return "image/jpeg" if p.endswith(".svg"): return "image/svg+xml" return "application/octet-stream" def commit_changes(changes): """POST update-file with the given changes; bumps revn cache.""" global _revn_cache if _revn_cache is None: _revn_cache = get_revn() payload = {"id": FILE_ID, "sessionId": str(uuid.uuid4()), "revn": _revn_cache, "vern": 0, "changes": changes} r = subprocess.run([ "curl", "-s", "-b", COOKIE, "-X", "POST", f"{BASE_URL}/api/rpc/command/update-file", "-H", "Content-Type: application/json", "-H", "Accept: application/json", "-d", json.dumps(payload), ], capture_output=True, text=True) try: d = json.loads(r.stdout) except json.JSONDecodeError: raise RuntimeError(f"non-JSON response: {r.stdout[:500]}") if isinstance(d, dict) and d.get("type") == "validation": raise RuntimeError(f"validation error: {r.stdout[:600]}") new_revn = d.get("revn", _revn_cache) if isinstance(d, dict) else _revn_cache _revn_cache = new_revn + 1 # next call will be a new revn print(f" Committed {len(changes)} changes (revn -> {_revn_cache})") return d # ───────────────────────────────────────────────────────────────────────────── # Shape constructors # ───────────────────────────────────────────────────────────────────────────── def _geom(x, y, w, h): return { "x": x, "y": y, "width": w, "height": h, "rotation": 0, "selrect": {"x": x, "y": y, "x1": x, "y1": y, "x2": x+w, "y2": y+h, "width": w, "height": h}, "points": [{"x": x, "y": y}, {"x": x+w, "y": y}, {"x": x+w, "y": y+h}, {"x": x, "y": y+h}], "transform": {"a":1,"b":0,"c":0,"d":1,"e":0,"f":0}, "transformInverse": {"a":1,"b":0,"c":0,"d":1,"e":0,"f":0}, } def frame(name, x, y, w, h, fill=BG_LIGHT, sid=None): sid = sid or str(uuid.uuid4()) obj = { "id": sid, "type": "frame", "name": name, **_geom(x, y, w, h), "parentId": ROOT, "frameId": ROOT, "fills": [{"fillColor": fill, "fillOpacity": 1}], "shapes": [], } return sid, obj def rect(name, x, y, w, h, parent_frame, fill=None, stroke=None, rx=0, sid=None): sid = sid or str(uuid.uuid4()) obj = {"id": sid, "type": "rect", "name": name, **_geom(x, y, w, h), "parentId": parent_frame, "frameId": parent_frame} obj["fills"] = [{"fillColor": fill, "fillOpacity": 1}] if fill else [] if stroke: sw, sc = stroke obj["strokes"] = [{"strokeColor": sc, "strokeOpacity": 1, "strokeStyle": "solid", "strokeWidth": sw, "strokeAlignment": "center"}] if rx: obj["rx"] = rx obj["ry"] = rx return sid, obj def text(name, x, y, w, h, content, parent_frame, size=14, weight="400", color=TEXT_DARK, align="left", sid=None): sid = sid or str(uuid.uuid4()) obj = { "id": sid, "type": "text", "name": name, **_geom(x, y, w, h), "parentId": parent_frame, "frameId": parent_frame, "growType": "fixed", "content": { "type": "root", "children": [{ "type": "paragraph-set", "children": [{ "type": "paragraph", "textAlign": align, "children": [{ "text": content, "fontId": "sourcesanspro", "fontFamily": "sourcesanspro", "fontSize": str(size), "fontWeight": weight, "fills": [{"fillColor": color, "fillOpacity": 1}], }], }], }], }, } return sid, obj def image(name, x, y, w, h, media_id, mw, mh, parent_frame, mtype="image/png", sid=None): sid = sid or str(uuid.uuid4()) obj = { "id": sid, "type": "image", "name": name, **_geom(x, y, w, h), "parentId": parent_frame, "frameId": parent_frame, "metadata": {"id": media_id, "width": mw, "height": mh, "mtype": mtype}, } return sid, obj def add_obj_change(obj, parent): return {"type": "add-obj", "id": obj["id"], "pageId": PAGE_ID, "frameId": parent, "parentId": parent, "obj": obj} # ───────────────────────────────────────────────────────────────────────────── # Figure builders # ───────────────────────────────────────────────────────────────────────────── def fig_y(figure_index, custom_heights=None): """Stable y-coordinate for figure N (1..6) given per-figure heights.""" if custom_heights is None: custom_heights = {1: 600, 2: 540, 3: 700, 4: 550, 5: 480, 6: 480} y = 100 for i in range(1, figure_index): y += custom_heights.get(i, 600) + FRAME_GAP return y # ───────────────────────────────────────────────────────────────────────────── # Figure 1: Overview teaser # ───────────────────────────────────────────────────────────────────────────── def build_fig1_overview(): fy = fig_y(1) fh = 600 print(f"\n=== Fig 1: Overview teaser ===") fid, fobj = frame("Fig1_Overview", FRAME_X, fy, FRAME_W, fh) changes = [add_obj_change(fobj, ROOT)] # Title _, t = text("fig1-title", FRAME_X+30, fy+15, FRAME_W-60, 32, "Figure 1: PARA — Pixel-Aligned Robot Actions", fid, size=22, weight="700") changes.append(add_obj_change(t, fid)) # ── 3 panels: left (diagram), middle (vignettes), right (numbers) ── panel_w = (FRAME_W - 90) // 3 # 30 + p1 + 15 + p2 + 15 + p3 + 30 panel_x = [FRAME_X+30, FRAME_X+30 + panel_w + 15, FRAME_X+30 + 2*(panel_w + 15)] panel_y = fy + 65 panel_h = fh - 100 # ── LEFT PANEL: Comparison diagram ── px = panel_x[0] _, lbl = text("p1-label", px, panel_y, panel_w, 22, "(a) Architecture Comparison", fid, size=13, weight="700") changes.append(add_obj_change(lbl, fid)) # Global Regression row grow_y = panel_y + 40 _, grow_lbl = text("p1-grow-lbl", px, grow_y, panel_w, 18, "Global Regression (ACT)", fid, size=11, weight="600", color=ACT_RED) changes.append(add_obj_change(grow_lbl, fid)) box_w, box_h = 60, 38 boxes_g = [("Image", "#eaf2fb", "#3b6fa6"), ("DINO", "#fff3e6", "#d97e1f"), ("CLS", "#f5e6fb", "#c2415a"), ("MLP", "#fff", "#71717a"), ("(x,y,z)", "#e8f5ec", ACT_RED)] spacing = (panel_w - box_w * len(boxes_g)) // (len(boxes_g) - 1) for i, (label, fill, stroke) in enumerate(boxes_g): bx = px + i * (box_w + spacing) _, b = rect(f"p1-grow-box-{i}", bx, grow_y + 22, box_w, box_h, fid, fill=fill, stroke=(1.5, stroke), rx=6) _, bt = text(f"p1-grow-box-{i}-t", bx, grow_y + 22 + (box_h-14)//2, box_w, 16, label, fid, size=10, weight="600", align="center") changes.extend([add_obj_change(b, fid), add_obj_change(bt, fid)]) # Arrow between boxes if i < len(boxes_g) - 1: ax = bx + box_w + 2 ay = grow_y + 22 + box_h // 2 - 1 _, a = rect(f"p1-grow-arr-{i}", ax, ay, spacing - 6, 2, fid, fill=NEUTRAL) changes.append(add_obj_change(a, fid)) # PARA row prow_y = grow_y + 95 _, prow_lbl = text("p1-prow-lbl", px, prow_y, panel_w, 18, "PARA", fid, size=11, weight="600", color=PARA_GREEN) changes.append(add_obj_change(prow_lbl, fid)) boxes_p = [("Image", "#eaf2fb", "#3b6fa6"), ("DINO", "#fff3e6", "#d97e1f"), ("Heatmap\nVolume", "#fde9ec", "#c2415a"), ("Argmax", "#fff", "#71717a"), ("3D Pt", "#e8f5ec", PARA_GREEN)] box_w2 = 60 spacing2 = (panel_w - box_w2 * len(boxes_p)) // (len(boxes_p) - 1) for i, (label, fill, stroke) in enumerate(boxes_p): bx = px + i * (box_w2 + spacing2) _, b = rect(f"p1-para-box-{i}", bx, prow_y + 22, box_w2, box_h, fid, fill=fill, stroke=(1.5, stroke), rx=6) # Multi-line shorter labels just use single line at smaller size first_line = label.split("\n")[0] if "\n" in label else label _, bt = text(f"p1-para-box-{i}-t", bx, prow_y + 22 + (box_h-14)//2, box_w2, 16, first_line, fid, size=10, weight="600", align="center") changes.extend([add_obj_change(b, fid), add_obj_change(bt, fid)]) if i < len(boxes_p) - 1: ax = bx + box_w2 + 2 ay = prow_y + 22 + box_h // 2 - 1 _, a = rect(f"p1-para-arr-{i}", ax, ay, spacing2 - 6, 2, fid, fill=NEUTRAL) changes.append(add_obj_change(a, fid)) # Punchline text below _, punch = text("p1-punch", px, prow_y + 100, panel_w, 32, "CLS collapses spatial info; PARA preserves it.", fid, size=11, weight="500", color=SUBTLE, align="left") changes.append(add_obj_change(punch, fid)) # ── MIDDLE PANEL: 3 vignettes ── px = panel_x[1] _, lbl2 = text("p2-label", px, panel_y, panel_w, 22, "(b) Where PARA Helps", fid, size=13, weight="700") changes.append(add_obj_change(lbl2, fid)) vignettes = [ ("OOD Generalization", "Shifted camera & objects", PARA_GREEN), ("Video Backbone", "SVD video → robot", "#3b6fa6"), ("Cross-Embodiment", "Arm-deleted point tracks", "#5a3da2"), ] vh = (panel_h - 50) // 3 for i, (title, sub, color) in enumerate(vignettes): vy = panel_y + 35 + i * (vh + 8) # Icon box (placeholder colored circle) _, ic = rect(f"p2-vig-{i}-icon", px+10, vy+12, 36, 36, fid, fill=color, rx=18) _, tt = text(f"p2-vig-{i}-title", px+58, vy+10, panel_w-70, 22, title, fid, size=12, weight="700") _, st = text(f"p2-vig-{i}-sub", px+58, vy+30, panel_w-70, 18, sub, fid, size=10, weight="400", color=SUBTLE) for x in (ic, tt, st): changes.append(add_obj_change(x, fid)) # ── RIGHT PANEL: 3 headline numbers ── px = panel_x[2] _, lbl3 = text("p3-label", px, panel_y, panel_w, 22, "(c) Headline Results", fid, size=13, weight="700") changes.append(add_obj_change(lbl3, fid)) headlines = [ ("97% vs 9%", "Real Robot (20 demos)", PARA_GREEN, ACT_RED), ("90% vs 0%", "Video Backbone", PARA_GREEN, ACT_RED), ("[TBD]", "Point Track Pretraining", NEUTRAL, NEUTRAL), ] hh = (panel_h - 50) // 3 for i, (big, sub, c1, c2) in enumerate(headlines): hy = panel_y + 35 + i * (hh + 8) # Background card _, card = rect(f"p3-card-{i}", px, hy, panel_w, hh, fid, fill="#ffffff", stroke=(1.5, BORDER), rx=10) _, big_t = text(f"p3-big-{i}", px, hy+12, panel_w, 36, big, fid, size=24, weight="800", color=c1, align="center") _, sub_t = text(f"p3-sub-{i}", px, hy+50, panel_w, 22, sub, fid, size=11, weight="500", color=SUBTLE, align="center") for x in (card, big_t, sub_t): changes.append(add_obj_change(x, fid)) commit_changes(changes) print(f" Done. Frame: {fid}") # ───────────────────────────────────────────────────────────────────────────── # Figure 2: Method details # ───────────────────────────────────────────────────────────────────────────── def build_fig2_method(): fy = fig_y(2) fh = 540 print(f"\n=== Fig 2: Method Details ===") fid, fobj = frame("Fig2_Method", FRAME_X, fy, FRAME_W, fh) changes = [add_obj_change(fobj, ROOT)] _, t = text("fig2-title", FRAME_X+30, fy+15, FRAME_W-60, 30, "Figure 2: PARA Method Pipeline", fid, size=20, weight="700") changes.append(add_obj_change(t, fid)) # Sub-labels _, la = text("fig2-a-label", FRAME_X+30, fy+60, 760, 22, "(a) Visualization of the PARA inference pipeline", fid, size=13, weight="700") _, lb = text("fig2-b-label", FRAME_X+820, fy+60, 360, 22, "(b) Height vs Depth", fid, size=13, weight="700") changes.extend([add_obj_change(la, fid), add_obj_change(lb, fid)]) # ── Panel (a): horizontal strip of 4 method frames ── strip_w = 760 strip_h = 360 sx = FRAME_X + 30 sy = fy + 95 # Upload all 4 frames method_frames = [] for i, fname in enumerate([ "frame_2_stage1.png", "frame_2_stage2.png", "frame_2_stage3.png", "frame_2_stage4.png" ]): mid, mw, mh, mtype = upload_media( f"/data/cameron/penpot/figures/extracted/{fname}", f"Method Stage {i+1}") method_frames.append((mid, mw, mh, mtype)) img_w = (strip_w - 30) // 4 # 4 images, 3 gaps of 10 gap = 10 captions = ["Camera Frustum", "Heatmap Volume", "Argmax → 3D", "Robot at Target"] for i, ((mid, mw, mh, mtype), cap) in enumerate(zip(method_frames, captions)): ix = sx + i * (img_w + gap) ih = int(img_w * mh / mw) _, im = image(f"fig2-a-img-{i}", ix, sy, img_w, ih, mid, mw, mh, fid, mtype=mtype) changes.append(add_obj_change(im, fid)) _, ct = text(f"fig2-a-cap-{i}", ix, sy + ih + 6, img_w, 18, cap, fid, size=11, weight="600", color=SUBTLE, align="center") changes.append(add_obj_change(ct, fid)) # Arrow between if i < 3: ax = ix + img_w + 1 ay = sy + ih // 2 - 1 _, ar = rect(f"fig2-a-arr-{i}", ax, ay, gap-2, 2, fid, fill=TEXT_DARK) changes.append(add_obj_change(ar, fid)) # ── Panel (b): height illustration ── bx = FRAME_X + 820 by = fy + 95 bw = 360 hi_id, hi_w, hi_h, hi_mt = upload_media( "/data/cameron/para/paper/figs/figma/height_illustration.png", "Height Illustration") bh = int(bw * hi_h / hi_w) if bh > 380: bh = 380 bw = int(bh * hi_w / hi_h) _, im_h = image("fig2-b-img", bx, by, bw, bh, hi_id, hi_w, hi_h, fid, mtype=hi_mt) changes.append(add_obj_change(im_h, fid)) _, cap_h = text("fig2-b-cap", FRAME_X+820, by + bh + 12, 360, 18, "Height is camera-invariant; depth is not.", fid, size=11, weight="500", color=SUBTLE, align="center") changes.append(add_obj_change(cap_h, fid)) commit_changes(changes) print(f" Done. Frame: {fid}") # ───────────────────────────────────────────────────────────────────────────── # Figure 3: Real robot results — single image import # ───────────────────────────────────────────────────────────────────────────── def build_fig3_real_robot(): fy = fig_y(3) fh = 700 print(f"\n=== Fig 3: Real Robot Results ===") fid, fobj = frame("Fig3_RealRobot", FRAME_X, fy, FRAME_W, fh) changes = [add_obj_change(fobj, ROOT)] _, t = text("fig3-title", FRAME_X+30, fy+15, FRAME_W-60, 30, "Figure 3: Real Robot Results (SO-100, 20 demos)", fid, size=20, weight="700") changes.append(add_obj_change(t, fid)) # Import the existing Figma asset pr_id, pr_w, pr_h, pr_mt = upload_media( "/data/cameron/para/paper/figs/figma/para_results.png", "Para Results") iw = FRAME_W - 60 ih = int(iw * pr_h / pr_w) if ih > fh - 90: ih = fh - 90 iw = int(ih * pr_w / pr_h) ix = FRAME_X + (FRAME_W - iw) // 2 iy = fy + 60 _, im = image("fig3-img", ix, iy, iw, ih, pr_id, pr_w, pr_h, fid, mtype=pr_mt) changes.append(add_obj_change(im, fid)) commit_changes(changes) print(f" Done. Frame: {fid}") # ───────────────────────────────────────────────────────────────────────────── # Figure 4: OOD Analysis (rebuild — full version with all 3 panels) # ───────────────────────────────────────────────────────────────────────────── def build_fig4_full(): fy = fig_y(4) fh = 550 print(f"\n=== Fig 4: OOD Analysis (full) ===") fid, fobj = frame("Fig4_OOD_Analysis", FRAME_X, fy, FRAME_W, fh) changes = [add_obj_change(fobj, ROOT)] _, t = text("fig4-title", FRAME_X+30, fy+15, FRAME_W-60, 30, "Figure 4: OOD Generalization Analysis", fid, size=20, weight="700") changes.append(add_obj_change(t, fid)) panel_w = 380 panel_gap = 20 px_a = FRAME_X + 30 px_b = px_a + panel_w + panel_gap px_c = px_b + panel_w + panel_gap py = fy + 65 # Sub-labels _, la = text("fig4-a-label", px_a, py, panel_w, 22, "(a) Spatial Extrapolation", fid, size=13, weight="700") _, lb = text("fig4-b-label", px_b, py, panel_w, 22, "(b) Per-Angle Viewpoint Robustness", fid, size=13, weight="700") _, lc = text("fig4-c-label", px_c, py, panel_w, 22, "(c) Qualitative Comparison", fid, size=13, weight="700") changes.extend([add_obj_change(x, fid) for x in (la, lb, lc)]) panel_y = py + 30 panel_inner_h = 360 # Panel (a): distribution plot + bar chart dist_id, dw, dh, dmt = upload_media( "/data/cameron/para/.agents/reports/project_site/media/exp3_leftright_distribution.png", "Train/Test Distribution") dist_w = panel_w dist_h = int(dist_w * dh / dw) if dist_h > 240: dist_h = 240 dist_w = int(dist_h * dw / dh) _, dim = image("fig4-a-dist", px_a + (panel_w-dist_w)//2, panel_y, dist_w, dist_h, dist_id, dw, dh, fid, mtype=dmt) changes.append(add_obj_change(dim, fid)) # Bar chart underneath: PARA 54% vs ACT 1% bar_y = panel_y + dist_h + 10 bar_h = 60 bar_w_unit = (panel_w - 40) // 2 # PARA bar (54%) para_bar_h = int(bar_h * 0.54) _, pb = rect("fig4-a-para-bar", px_a + 20, bar_y + (bar_h - para_bar_h), (bar_w_unit-10), para_bar_h, fid, fill=PARA_GREEN) _, pbl = text("fig4-a-para-lbl", px_a + 20, bar_y + bar_h + 4, bar_w_unit-10, 18, "PARA 54%", fid, size=11, weight="700", color=PARA_GREEN, align="center") # ACT bar (1%) act_bar_h = max(2, int(bar_h * 0.01)) _, ab = rect("fig4-a-act-bar", px_a + 20 + bar_w_unit, bar_y + (bar_h - act_bar_h), (bar_w_unit-10), act_bar_h, fid, fill=ACT_RED) _, abl = text("fig4-a-act-lbl", px_a + 20 + bar_w_unit, bar_y + bar_h + 4, bar_w_unit-10, 18, "ACT 1%", fid, size=11, weight="700", color=ACT_RED, align="center") for x in (pb, pbl, ab, abl): changes.append(add_obj_change(x, fid)) # Panel (b): per-theta chart pt_id, pw, ph_, pmt = upload_media( "/data/cameron/penpot/figures/per_theta.png", "Per-Theta Chart (2)") chart_w = panel_w chart_h = int(chart_w * ph_ / pw) if chart_h > panel_inner_h: chart_h = panel_inner_h chart_w = int(chart_h * pw / ph_) _, ci = image("fig4-b-chart", px_b + (panel_w-chart_w)//2, panel_y, chart_w, chart_h, pt_id, pw, ph_, fid, mtype=pmt) changes.append(add_obj_change(ci, fid)) # Panel (c): 2x2 grid of feature comparison frames cell_w = (panel_w - 10) // 2 cell_h = cell_w cells = [ ("4c_act_start.png", px_c, panel_y, "ACT start"), ("4c_act_mid.png", px_c+cell_w+10, panel_y, "ACT mid"), ("4c_para_start.png", px_c, panel_y+cell_h+24,"PARA start"), ("4c_para_mid.png", px_c+cell_w+10, panel_y+cell_h+24,"PARA mid"), ] for i, (fn, cx, cy, label) in enumerate(cells): mid, mw, mh, mtype = upload_media( f"/data/cameron/penpot/figures/extracted/{fn}", f"Fig4c {label}") _, im = image(f"fig4-c-{i}", cx, cy, cell_w, cell_h, mid, mw, mh, fid, mtype=mtype) _, cap = text(f"fig4-c-cap-{i}", cx, cy + cell_h + 2, cell_w, 18, label, fid, size=10, weight="600", color=SUBTLE, align="center") changes.extend([add_obj_change(im, fid), add_obj_change(cap, fid)]) # Verdict text under panel c _, vd = text("fig4-c-verdict", px_c, panel_y + 2*(cell_h+24) + 10, panel_w, 18, "ACT reaches wrong location · PARA reaches bowl correctly", fid, size=10, weight="500", color=SUBTLE, align="center") changes.append(add_obj_change(vd, fid)) commit_changes(changes) print(f" Done. Frame: {fid}") # ───────────────────────────────────────────────────────────────────────────── # Figure 5: Video Backbone # ───────────────────────────────────────────────────────────────────────────── def build_fig5_video_backbone(): fy = fig_y(5) fh = 480 print(f"\n=== Fig 5: Video Backbone ===") fid, fobj = frame("Fig5_VideoBackbone", FRAME_X, fy, FRAME_W, fh) changes = [add_obj_change(fobj, ROOT)] _, t = text("fig5-title", FRAME_X+30, fy+15, FRAME_W-60, 30, "Figure 5: Video Models as Policy Backbones", fid, size=20, weight="700") changes.append(add_obj_change(t, fid)) # Left: training diagram diag_x = FRAME_X + 30 diag_y = fy + 70 diag_w = 480 _, lab = text("fig5-diag-label", diag_x, diag_y, diag_w, 22, "(a) Two-Stage Training", fid, size=13, weight="700") changes.append(add_obj_change(lab, fid)) boxes = [ ("Video UNet\n(4K pretrain)", diag_x + 20, diag_y + 100, 160, 72, "#eaf2fb", "#3b6fa6"), ("Joint Fine-tune\n(3K)", diag_x + 240, diag_y + 100, 160, 72, "#fff3e6", "#d97e1f"), ] for name, x, y, w, h, fill, stroke in boxes: _, b = rect(f"fig5-{name[:8]}", x, y, w, h, fid, fill=fill, stroke=(2, stroke), rx=10) first_line = name.split("\n")[0] sub = name.split("\n")[1] if "\n" in name else "" _, bt = text(f"fig5-{name[:8]}-t", x, y+18, w, 22, first_line, fid, size=13, weight="700", align="center") _, bs = text(f"fig5-{name[:8]}-s", x, y+40, w, 18, sub, fid, size=10, weight="400", color=SUBTLE, align="center") changes.extend([add_obj_change(b, fid), add_obj_change(bt, fid), add_obj_change(bs, fid)]) # Arrow between _, ar = rect("fig5-arr", diag_x + 180, diag_y + 134, 60, 2, fid, fill=TEXT_DARK) changes.append(add_obj_change(ar, fid)) # Two output branches from joint fine-tune out_y_top = diag_y + 90 out_y_bot = diag_y + 200 _, out1 = text("fig5-out1", diag_x + 240, out_y_top - 20, 160, 16, "↑ PARA Heatmap Head", fid, size=10, weight="600", color=PARA_GREEN, align="center") _, out2 = text("fig5-out2", diag_x + 240, out_y_bot - 5, 160, 16, "↓ Video Generation", fid, size=10, weight="600", color=SUBTLE, align="center") changes.extend([add_obj_change(out1, fid), add_obj_change(out2, fid)]) # Right: rollout grids comparison right_x = FRAME_X + 540 right_y = fy + 70 right_w = 620 _, rlab = text("fig5-right-label", right_x, right_y, right_w, 22, "(b) Rollout Comparison", fid, size=13, weight="700") changes.append(add_obj_change(rlab, fid)) para_id, pw, ph_, pmt = upload_media( "/data/cameron/penpot/figures/extracted/frame_5_para.png", "Rollout PARA") glob_id, gw, gh, gmt = upload_media( "/data/cameron/penpot/figures/extracted/frame_5_global.png", "Rollout Global") cell_w = (right_w - 20) // 2 cell_h = int(cell_w * ph_ / pw) if cell_h > 280: cell_h = 280 _, im_p = image("fig5-rollout-para", right_x, right_y + 30, cell_w, cell_h, para_id, pw, ph_, fid, mtype=pmt) _, im_g = image("fig5-rollout-global", right_x + cell_w + 20, right_y + 30, cell_w, cell_h, glob_id, gw, gh, fid, mtype=gmt) changes.extend([add_obj_change(im_p, fid), add_obj_change(im_g, fid)]) _, cp = text("fig5-rollout-para-cap", right_x, right_y + 30 + cell_h + 6, cell_w, 22, "SVD + PARA — 92%", fid, size=14, weight="700", color=PARA_GREEN, align="center") _, cg = text("fig5-rollout-global-cap", right_x + cell_w + 20, right_y + 30 + cell_h + 6, cell_w, 22, "SVD + Global Regression — 0%", fid, size=14, weight="700", color=ACT_RED, align="center") changes.extend([add_obj_change(cp, fid), add_obj_change(cg, fid)]) commit_changes(changes) print(f" Done. Frame: {fid}") # ───────────────────────────────────────────────────────────────────────────── # Figure 6: Point Track Pretraining (placeholder) # ───────────────────────────────────────────────────────────────────────────── def build_fig6_point_tracks(): fy = fig_y(6) fh = 480 print(f"\n=== Fig 6: Point Track Pretraining (placeholder) ===") fid, fobj = frame("Fig6_PointTracks", FRAME_X, fy, FRAME_W, fh) changes = [add_obj_change(fobj, ROOT)] _, t = text("fig6-title", FRAME_X+30, fy+15, FRAME_W-60, 30, "Figure 6: Point Track Pretraining (Preliminary)", fid, size=20, weight="700") changes.append(add_obj_change(t, fid)) # Left: 3 placeholder frames left_x = FRAME_X + 30 left_y = fy + 70 left_w = 540 _, lab = text("fig6-left-label", left_x, left_y, left_w, 22, "(a) Arm-Deleted Training Data", fid, size=13, weight="700") changes.append(add_obj_change(lab, fid)) cell_w = (left_w - 20) // 3 cell_h = cell_w * 3 // 4 table_color = "#e8d5b7" for i in range(3): cx = left_x + i * (cell_w + 10) cy = left_y + 35 # Placeholder: tan rectangle (table) with a small green dot (EEF) _, table = rect(f"fig6-table-{i}", cx, cy, cell_w, cell_h, fid, fill=table_color, stroke=(1, BORDER), rx=4) # Object (small circle stand-in) ox = cx + cell_w//2 - 8 oy = cy + cell_h//2 _, obj = rect(f"fig6-obj-{i}", ox, oy, 16, 8, fid, fill="#888888", rx=4) # Green dot (EEF position) dx = cx + cell_w//2 + 20 + i*4 dy = cy + cell_h//2 - 12 _, dot = rect(f"fig6-dot-{i}", dx, dy, 8, 8, fid, fill=PARA_GREEN, rx=4) # Frame number _, fn = text(f"fig6-fn-{i}", cx, cy + cell_h + 2, cell_w, 16, f"t = {i*5}", fid, size=10, color=SUBTLE, align="center") for x in (table, obj, dot, fn): changes.append(add_obj_change(x, fid)) _, leftcap = text("fig6-leftcap", left_x, left_y + 35 + cell_h + 24, left_w, 18, "(robot arm invisible; only EEF position retained as supervision)", fid, size=10, weight="500", color=SUBTLE, align="center") changes.append(add_obj_change(leftcap, fid)) # Right: placeholder bar chart right_x = FRAME_X + 600 right_y = fy + 70 right_w = 560 _, rlab = text("fig6-right-label", right_x, right_y, right_w, 22, "(b) Pretrain → Fine-tune Results (preliminary)", fid, size=13, weight="700") changes.append(add_obj_change(rlab, fid)) # 4 bars: PARA pretrain, PARA scratch, Global pretrain, Global scratch bar_data = [ ("PARA\npretrain", 70, PARA_GREEN), ("PARA\nscratch", 30, PARA_GREEN), ("Global\npretrain", 20, ACT_RED), ("Global\nscratch", 15, ACT_RED), ] plot_x = right_x + 30 plot_y = right_y + 60 plot_w = right_w - 60 plot_h = 240 bar_w = (plot_w - 40) // 4 bar_gap = (plot_w - bar_w*4) // 5 # Y-axis line _, ya = rect("fig6-y-axis", plot_x, plot_y, 2, plot_h, fid, fill="#888888") # X-axis line _, xa = rect("fig6-x-axis", plot_x, plot_y + plot_h - 1, plot_w, 2, fid, fill="#888888") changes.extend([add_obj_change(ya, fid), add_obj_change(xa, fid)]) for i, (label, val, color) in enumerate(bar_data): bh = int(plot_h * val / 100) bx = plot_x + bar_gap + i * (bar_w + bar_gap) by = plot_y + plot_h - bh _, b = rect(f"fig6-bar-{i}", bx, by, bar_w, bh, fid, fill=color) # Hatching effect for "scratch" — overlay with lower opacity color (using lighter color rect) _, bv = text(f"fig6-bv-{i}", bx, by - 18, bar_w, 16, f"{val}%", fid, size=11, weight="700", align="center") _, bl = text(f"fig6-bl-{i}", bx-4, plot_y + plot_h + 4, bar_w+8, 28, label.replace("\n", " "), fid, size=9, weight="500", color=SUBTLE, align="center") for x in (b, bv, bl): changes.append(add_obj_change(x, fid)) _, rcap = text("fig6-rcap", right_x, right_y + plot_h + 80, right_w, 18, "Preliminary — final values pending from backbones agent", fid, size=10, weight="600", color=NEUTRAL, align="center") changes.append(add_obj_change(rcap, fid)) commit_changes(changes) print(f" Done. Frame: {fid}") def build_fig4_with_chart_only(): """Fig 4: OOD Generalization Analysis. For now: just panel (b) per-theta chart. Full figure (with 4a distribution + 4c qualitative comparison) added later. """ fy = fig_y(4) fh = 700 print(f"\n=== Fig 4: OOD Analysis (placing panel 4b first) ===") print(f" Frame at ({FRAME_X},{fy}) {FRAME_W}x{fh}") fid, fobj = frame("Fig4_OOD_Analysis", FRAME_X, fy, FRAME_W, fh) changes = [add_obj_change(fobj, ROOT)] # Title _, t = text("fig4-title", FRAME_X + 30, fy + 20, FRAME_W - 60, 30, "Figure 4: OOD Generalization Analysis", fid, size=20, weight="700", align="left") changes.append(add_obj_change(t, fid)) # Panel labels _, lab_a = text("panel-a-label", FRAME_X + 30, fy + 70, 380, 22, "(a) Spatial Extrapolation", fid, size=14, weight="700", align="left") _, lab_b = text("panel-b-label", FRAME_X + 430, fy + 70, 380, 22, "(b) Per-Angle Viewpoint Robustness", fid, size=14, weight="700", align="left") _, lab_c = text("panel-c-label", FRAME_X + 830, fy + 70, 380, 22, "(c) Qualitative Comparison", fid, size=14, weight="700", align="left") for lab in (lab_a, lab_b, lab_c): changes.append(add_obj_change(lab, fid)) # Upload + place per-theta chart in panel (b) media_id, mw, mh, mtype = upload_media( "/data/cameron/penpot/figures/per_theta.png", "Per-Theta Chart") chart_w = 380 chart_h = int(chart_w * mh / mw) chart_x = FRAME_X + 430 chart_y = fy + 110 _, img = image("per-theta-chart", chart_x, chart_y, chart_w, chart_h, media_id, mw, mh, fid, mtype=mtype) changes.append(add_obj_change(img, fid)) # Placeholders for 4a and 4c (to be filled in later) _, ph_a = rect("placeholder-4a", FRAME_X + 30, fy + 110, 380, 280, fid, stroke=(2, BORDER), rx=8) _, ph_c = rect("placeholder-4c", FRAME_X + 830, fy + 110, 380, 280, fid, stroke=(2, BORDER), rx=8) changes.append(add_obj_change(ph_a, fid)) changes.append(add_obj_change(ph_c, fid)) _, ph_a_t = text("placeholder-4a-text", FRAME_X + 30, fy + 240, 380, 30, "(distribution plot — pending)", fid, size=12, color=NEUTRAL, align="center") _, ph_c_t = text("placeholder-4c-text", FRAME_X + 830, fy + 240, 380, 30, "(rollout frames — pending)", fid, size=12, color=NEUTRAL, align="center") changes.append(add_obj_change(ph_a_t, fid)) changes.append(add_obj_change(ph_c_t, fid)) commit_changes(changes) print(f" Done. Frame: {fid}") # ───────────────────────────────────────────────────────────────────────────── # Main # ───────────────────────────────────────────────────────────────────────────── BUILDERS = { "1": build_fig1_overview, "2": build_fig2_method, "3": build_fig3_real_robot, "4": build_fig4_full, "4b": build_fig4_with_chart_only, "5": build_fig5_video_backbone, "6": build_fig6_point_tracks, } def main(): parser = argparse.ArgumentParser() parser.add_argument("--figs", default="4b", help="comma-separated list, or 'all'") args = parser.parse_args() if args.figs == "all": keys = list(BUILDERS.keys()) else: keys = [k.strip() for k in args.figs.split(",")] for k in keys: if k not in BUILDERS: print(f"unknown figure: {k} (have: {list(BUILDERS.keys())})") sys.exit(1) BUILDERS[k]() if __name__ == "__main__": main()