| """Parallel-investigation layout for the recorded study methodology.""" | |
| import numpy as np | |
| import matplotlib.pyplot as plt | |
| from matplotlib.patches import Rectangle, Ellipse, FancyArrowPatch | |
| from PIL import Image, ImageDraw, ImageFont, features | |
| INK='#233442' | |
| STAGES=['#386A9D','#A47832','#3B7960','#7A6398','#4D7E87','#247A78'] | |
| LANG_COLORS={'english':'#386A9D','hindi':'#C47D22','hinglish':'#6B6F91'} | |
| LABELS={'english':'English','hindi':'Hindi','hinglish':'Roman Hinglish'} | |
| def draw_methodology(base,scores,execution,example,detail): | |
| height=18.9 | |
| fig=plt.figure(figsize=(16,height),facecolor='white',dpi=120) | |
| figure_y=lambda top:1-top/height | |
| fig.text(.03,figure_y(.2025),'Study methodology',fontsize=18,weight='bold',color=INK,va='top') | |
| # Inch-based vertical geometry preserves panels 1–3 at their original size. | |
| specifications=[(.03,.405,.94,2.025),(.03,3.105,.435,2.43),(.535,3.105,.435,2.43), | |
| (.03,6.1425,.94,4.3),(.03,10.8425,.94,3.7),(.03,14.9425,.94,3.7)] | |
| bounds=[(x,figure_y(top+h),w,h/height) for x,top,w,h in specifications] | |
| titles=['Research setting','Pretrained VLM','Adapted VLA', | |
| 'Existing Evaluation Gaps', | |
| 'Lexical Sensitivity under Frozen-Policy Inference', | |
| 'Scene-Grounded Lexical Swapping'] | |
| axes=[] | |
| for i,(rect,title,color) in enumerate(zip(bounds,titles,STAGES)): | |
| ax=fig.add_axes(rect);ax.set(xlim=(0,1),ylim=(0,1));ax.axis('off') | |
| ax.add_patch(Rectangle((0,0),1,1,facecolor='white',edgecolor=color,linewidth=1.1)) | |
| header=.4725/(height*rect[3]);center=1-header/2 | |
| ax.add_patch(Rectangle((0,1-header),1,header,facecolor=color,alpha=.10,edgecolor='none')) | |
| cx=.30/(16*rect[2]);dx=.28/(16*rect[2]);dy=.28/(height*rect[3]) | |
| ax.add_patch(Ellipse((cx,center),dx,dy,facecolor=color,edgecolor='none')) | |
| ax.text(cx,center,str(i+1),ha='center',va='center',fontsize=12.5,weight='bold',color='white') | |
| ax.text(.62/(16*rect[2]),center,title,va='center',fontsize=15,weight='bold',color=INK) | |
| axes.append(ax) | |
| def text(ax,x,y,s,size=13,bold=False,color=INK,**kw): | |
| return ax.text(x,y,s,fontsize=size,weight='bold' if bold else 'normal',color=color,va='top',**kw) | |
| def hindi(ax,s,x,top,width,image_height,size=13): | |
| assert features.check('raqm'),'Pillow RAQM is required for the exact Hindi label' | |
| rect=ax.get_position();dpi=300 | |
| pxw=round(16*rect.width*width*dpi);pxh=round(height*rect.height*image_height*dpi) | |
| font=ImageFont.truetype(str(base/'notebook_backend/fonts/NotoSansDevanagari.ttf'),round(size*dpi/72),layout_engine=ImageFont.Layout.RAQM) | |
| image=Image.new('RGBA',(pxw,pxh),(255,255,255,0));draw=ImageDraw.Draw(image) | |
| box=draw.textbbox((0,0),s,font=font,language='hi',anchor='lt') | |
| assert box[2]<=pxw and box[3]<=pxh,(box,(pxw,pxh)) | |
| draw.text((0,0),s,font=font,fill=INK,language='hi',anchor='lt') | |
| ax.imshow(np.asarray(image),extent=(x,x+width,top-image_height,top),aspect='auto',interpolation='lanczos',zorder=2) | |
| def put(ax,x,top,s,size=14,bold=False,color=INK,**kw): | |
| panel_height=ax.get_position().height*height | |
| return text(ax,x,1-top/panel_height,s,size,bold,color,**kw) | |
| # Research context only: robot matching belongs exclusively to branch 3. | |
| ax=axes[0] | |
| ax.imshow(plt.imread(base/'data/backbone/libero_images/task8_state20_agentview_image.png'), | |
| extent=(.035,.1415,.055,.797),aspect='auto',interpolation='lanczos',zorder=2) | |
| text(ax,.175,.69,'LIBERO-Goal',14.5,True,color=STAGES[0]) | |
| text(ax,.175,.48,'10 manipulation tasks',12.6) | |
| text(ax,.175,.26,'Example: bowl on plate',12.6) | |
| for lang,x in [('english',.415),('hindi',.615),('hinglish',.805)]: | |
| text(ax,x,.68,LABELS[lang],13,True,color=LANG_COLORS[lang]) | |
| if lang=='hindi':hindi(ax,example[lang],x,.455,.18,.17) | |
| else: | |
| instruction=example[lang].replace(' on the plate','\non the plate').replace(' par rakho.','\npar rakho.') | |
| assert ' '.join(instruction.split())==example[lang] | |
| text(ax,x,.45,instruction,13) | |
| # Independent evaluation of the original model; numbers are secondary labels. | |
| ax=axes[1] | |
| text(ax,.045,.815,'Original Qwen3-VL-4B-Instruct',13.5,True,color=STAGES[1]) | |
| text(ax,.045,.688,'IndicVisionBench: 98 matched items / language',12.6) | |
| for lang,x in [('english',.045),('hindi',.35),('hinglish',.64)]: | |
| k,n=scores[lang];label='Hinglish' if lang=='hinglish' else LABELS[lang] | |
| text(ax,x,.578,f'{label} {k}/{n}',12,color=LANG_COLORS[lang]) | |
| text(ax,.045,.455,'LIBERO: text-only + image-conditioned probes',12.6) | |
| text(ax,.045,.344,'Numbered markers → corrected pink outlines',12.4) | |
| text(ax,.045,.205,'Multilingual instruction understanding',13,True,color=STAGES[1]) | |
| text(ax,.045,.095,'Scene recognition remained inconsistent',12.3) | |
| # Matched task IDs, initial states and policy apply to this branch only. | |
| ax=axes[2] | |
| text(ax,.045,.815,'Frozen StarVLA · LIBERO-Goal',13.5,True,color=STAGES[2]) | |
| text(ax,.045,.688,'Same task IDs + initial states + adapted policy',12.6) | |
| text(ax,.045,.575,'200 episodes / language · 600 matched outcomes',12.4) | |
| for lang,y in [('english',.415),('hindi',.275),('hinglish',.135)]: | |
| k,n=execution[lang] | |
| text(ax,.045,y+.029,LABELS[lang],12.1,True,color=LANG_COLORS[lang]) | |
| ax.add_patch(Rectangle((.345,y-.027),.39,.055,facecolor='#edf0f3',edgecolor='none')) | |
| ax.add_patch(Rectangle((.345,y-.027),.39*k/n,.055,facecolor=LANG_COLORS[lang],edgecolor='none')) | |
| text(ax,.765,y+.029,f'{k}/{n}',12.4) | |
| # Two complementary investigations, explained with the actual audit example. | |
| ax=axes[3] | |
| put(ax,.025,.64,'Existing robot benchmarks',14.5,True,color=STAGES[3]) | |
| put(ax,.025,.98,'Standard robot benchmarks',14,True) | |
| put(ax,.025,1.27,'LIBERO, CALVIN, and RoboCasa evaluate whether robots\n' | |
| 'can successfully execute task instructions, primarily in English.') | |
| put(ax,.025,1.94,'Instruction robustness',14,True) | |
| put(ax,.025,2.23,'LIBERO-Plus and LIBERO-Para investigate whether robots\n' | |
| 'still perform the same task when instructions are reworded\n' | |
| 'or their descriptions change.') | |
| put(ax,.025,3.12,'What remains unexplored',14,True,color='#A65353') | |
| put(ax,.025,3.41,'Natural Hindi and Roman Hinglish instruction following is\n' | |
| 'not directly evaluated in these original protocols.') | |
| ax.plot([.50,.50],[.045,.865],color='#d6dce1',lw=.8) | |
| put(ax,.525,.64,'Checking Machine Translations',14.5,True,color=STAGES[3]) | |
| put(ax,.525,1.02,'10 original English LIBERO instructions were translated\n' | |
| 'into Hindi and French using Google Translate, DeepL,\nand Yandex.') | |
| put(ax,.525,1.91,'The 60 resulting translations were inspected for preservation\n' | |
| 'of the original objects, actions, and spatial relationships.') | |
| put(ax,.525,2.55,'Back-translation into English was also examined to inspect\n' | |
| 'changes in meaning.') | |
| put(ax,.525,3.12,'Original English: “Push the plate”',14,True) | |
| put(ax,.525,3.39,'Yandex Hindi:',14) | |
| hindi(ax,'दबाएं',.64,1-3.36/4.3,.065,.085,size=14) | |
| put(ax,.705,3.39,'(press)',14) | |
| put(ax,.525,3.66,'Changed meaning: “Press the plate”',14,True,color='#A65353') | |
| put(ax,.525,4.00,'Translation can change the physical action requested of the robot.',13.4,True,color=STAGES[3]) | |
| # The same intended goal, with only the substituted destination noun emphasized. | |
| ax=axes[4] | |
| put(ax,.025,.61,'Bowl on cabinet · destination-noun substitution · task success',14,True) | |
| for x,label,phrase,count,color in [ | |
| (.025,'Original French',"l'armoire.",'0/5','#A65353'), | |
| (.525,'Modified instruction','the cabinet.','3/5',STAGES[2])]: | |
| put(ax,x,.95,label,14,True,color=STAGES[4]) | |
| put(ax,x+.405,.94,count,24,True,color=color) | |
| put(ax,x,1.35,'Placez le bol sur le dessus de',14) | |
| put(ax,x+.205,1.35,phrase,14,True,color=color) | |
| put(ax,x,1.70,'English meaning: “Place the bowl on top of the cabinet.”',13.2, | |
| color='#596b78',fontstyle='italic') | |
| ax.plot([.025,.975],[1-2.02/3.7]*2,color='#d6dce1',lw=.8) | |
| put(ax,.025,2.16,'Instruction structure · 20 recorded episodes',13.5,True,color=STAGES[4]) | |
| for x,label,count,criterion,example_s in [ | |
| (.025,'Keywords only','2/4','Either monitored destination','Bowl cabinet plate'), | |
| (.35,'Target-first','7/8','Requested destination first', | |
| 'Bowl ko cabinet ke upar rakho,\nplate par nahi.'), | |
| (.675,'Negation-first / reordered','4/8','Requested destination first', | |
| 'Plate par nahi, bowl ko\ncabinet ke upar rakho.')]: | |
| put(ax,x,2.50,label+' · '+count,13,True) | |
| put(ax,x,2.78,criterion,12.3,color='#596b78') | |
| put(ax,x,3.10,example_s,12.3) | |
| # Proposed input intervention, visually distinct from the measured outcomes. | |
| ax=axes[5] | |
| nodes=[(.025,.245),(.30,.325),(.655,.14),(.815,.16)] | |
| for x,width in nodes: | |
| ax.add_patch(Rectangle((x,.04),width,1-.60/3.7-.04, | |
| facecolor='#eff7f6' if x<.65 else '#edf4ef', | |
| edgecolor=STAGES[5],linewidth=.9,linestyle='--' if x==.30 else '-')) | |
| put(ax,.035,.73,'Natural multilingual instructions',13.2,True,color=STAGES[5]) | |
| put(ax,.035,1.06,'Hindi',13.4,True,color=LANG_COLORS['hindi']) | |
| hindi(ax,'कैबिनेट की बीच की दराज खोलें',.035,1-1.36/3.7,.225,.105,size=14) | |
| put(ax,.035,1.75,'English: Open the middle drawer\nof the cabinet.',13.2, | |
| color='#596b78',fontstyle='italic') | |
| put(ax,.035,2.38,'Roman Hinglish',13.4,True,color=LANG_COLORS['hinglish']) | |
| put(ax,.035,2.70,'cabinet ka beech wala drawer kholo',13.5) | |
| put(ax,.035,3.05,'English: Open the middle drawer\nof the cabinet.',13.2, | |
| color='#596b78',fontstyle='italic') | |
| put(ax,.31,.73,'Grounded lexical swapping',13.5,True,color=STAGES[5]) | |
| put(ax,.31,1.10,'Identify the intended task from the scene,\n' | |
| 'then replace selected terms with\npolicy-familiar vocabulary.',13.5) | |
| put(ax,.31,2.04,'Preserve the intended action, objects,\n' | |
| 'destinations, and spatial relationships.',13.5) | |
| put(ax,.31,2.83,'Hindi → cabinet ka middle drawer kholo',13.4) | |
| put(ax,.31,3.21,'Roman Hinglish → cabinet ka middle drawer kholo',13.4) | |
| put(ax,.665,.73,'Existing policy',13.5,True,color=STAGES[5]) | |
| put(ax,.665,1.65,'StarVLA ·\nQwen3-VL + OFT',13.5) | |
| put(ax,.665,2.55,'Same trained\nweights',13.2) | |
| put(ax,.825,.73,'Expected improvement',12.5,color=STAGES[5]) | |
| put(ax,.825,1.65,'Higher multilingual\ntask success',14,True,color=STAGES[5]) | |
| for (x,width),(next_x,_) in zip(nodes,nodes[1:]): | |
| ax.add_patch(FancyArrowPatch((x+width+.004,.47),(next_x-.004,.47),arrowstyle='-|>', | |
| mutation_scale=12,color=STAGES[5],linewidth=1.2,shrinkA=0,shrinkB=0)) | |
| # Parallel evidence branches converge; complementary auditing belongs to stage 4. | |
| color='#788995' | |
| def line(points): | |
| fig.add_artist(plt.Line2D([p[0] for p in points],[p[1] for p in points],transform=fig.transFigure,color=color,linewidth=1.25)) | |
| def arrow(a,b): | |
| fig.add_artist(FancyArrowPatch(a,b,transform=fig.transFigure,arrowstyle='-|>',mutation_scale=14,color=color,linewidth=1.25,shrinkA=0,shrinkB=0)) | |
| original_y=lambda value:figure_y((1-value)*13.5) | |
| def original_line(points):line([(x,original_y(v)) for x,v in points]) | |
| def original_arrow(a,b):arrow((a[0],original_y(a[1])),(b[0],original_y(b[1]))) | |
| original_line([(.5,.814),(.5,.79),(.2475,.79)]) | |
| original_line([(.5,.79),(.7525,.79)]) | |
| original_arrow((.2475,.79),(.2475,.776));original_arrow((.7525,.79),(.7525,.776)) | |
| original_line([(.2475,.584),(.2475,.56),(.7525,.56)]) | |
| original_line([(.7525,.584),(.7525,.56)]) | |
| original_arrow((.7525,.56),(.7525,.551)) | |
| arrow((.5,figure_y(10.52)),(.5,figure_y(10.76))) | |
| arrow((.5,figure_y(14.62)),(.5,figure_y(14.86))) | |
| fig.canvas.draw();renderer=fig.canvas.get_renderer();overflow=[] | |
| for i,ax in enumerate(axes): | |
| box=ax.get_window_extent(renderer) | |
| for artist in ax.texts: | |
| tbox=artist.get_window_extent(renderer) | |
| if tbox.x0<box.x0-1 or tbox.x1>box.x1+1 or tbox.y0<box.y0-1 or tbox.y1>box.y1+1: | |
| overflow.append((i+1,artist.get_text())) | |
| for i,ax in enumerate(axes[3:],start=4): | |
| for j,first in enumerate(ax.texts): | |
| for second in ax.texts[j+1:]: | |
| if first.get_window_extent(renderer).overlaps(second.get_window_extent(renderer)): | |
| overflow.append((i,'Overlapping text: '+first.get_text()+' / '+second.get_text())) | |
| for artist in axes[5].texts[2:]: | |
| x=artist.get_position()[0] | |
| node=next(((left,width) for left,width in nodes if left<=x<left+width),None) | |
| if node: | |
| left,width=node | |
| minimum=axes[5].transData.transform((left,0))[0] | |
| maximum=axes[5].transData.transform((left+width,0))[0] | |
| box=artist.get_window_extent(renderer) | |
| if box.x0<minimum-1 or box.x1>maximum+1: | |
| overflow.append((6,'Text exceeds block: '+artist.get_text())) | |
| assert not overflow,overflow | |
| return fig,{'text_overflow':overflow,'figure_inches':[16,height], | |
| 'panel_geometry':[{'x_fraction':x,'top_inches':top,'width_fraction':w,'height_inches':h} | |
| for x,top,w,h in specifications], | |
| 'panels_1_to_3_geometry_preserved':True, | |
| 'connections':[[1,2],[1,3],[2,4],[3,4],['translation_audit',4],[4,5],[5,6]], | |
| 'finding':'Multilingual instruction understanding', | |
| 'scene_recognition':'inconsistent','robot_matching_stage':3} | |
Xet Storage Details
- Size:
- 13.9 kB
- Xet hash:
- 74bf85cad7480a14f6e537d73134e8396f60001dc81382a310605494a5922c18
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.