diff --git a/_includes/robot-policy/act-pipeline.svg b/_includes/robot-policy/act-pipeline.svg index fd4d7e56..068f6118 100644 --- a/_includes/robot-policy/act-pipeline.svg +++ b/_includes/robot-policy/act-pipeline.svg @@ -1,324 +1 @@ - - ACT inference pipeline on the NexArm - Three camera frames and the six joint angles go in. A ResNet18 turns each frame into 300 tokens; with a state token and a style-latent token that is zero at inference they form 902 tokens. Four encoder layers mix them, one decoder layer with 100 query slots reads the result, and a linear layer gives 100 joint targets, all of which the arm runs before it looks again. - - - - - - - ACT turns one look into 100 actions. - - - - - - - - - - - 1 · observe - - - - Cameras and joints - - FRONT - - WRIST - - TOP - - j1 - - j2 - - j3 - - j4 - - j5 - - grip - z = 0 - STYLE LATENT - 3 × 480×640 RGB · state 6 - - - 2 · image backbone - - - - ResNet18 per camera - - 64 - - 64 - - 128 - - 256 - - 512 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 15 × 20 GRID - 512 × 15 × 20 → 300 tokens - - - 3 · encoder input - - - - One token sequence - - z​ - - s - - front 300 - - wrist 300 - - top 300 - - 902 TOKENS - (2 + 300 V) × 512 - - - 4 · transformer encoder - - - - Mix every token - - - - - self-attention - → feed-forward - × 4 - each token - sees all 902 - 4 layers · d 512 · 8 heads - - - 5 · decoder and action head - - - - 100 slots, 100 actions - - - - - - - - - - - - - - - - - - - - - - - - - - 100 QUERY SLOTS - - - cross-attention to tokens - - - - - - - - - - - - - - - - - - - - - - - - - - 100 × 6 - 1 layer → Linear 512 → 6 - - - 6 · execute - - - - Run all 100, then look - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 0 s - 3.33 s, then look again - 30 Hz · no re-planning in between - - source · lerobot/policies/act/modeling_act.py · click a block or use the arrow keys - +ACT inference pipeline on the NexArmThree camera frames and the six joint angles go in. A ResNet18 turns each frame into 300 tokens; with a state token and a style-latent token that is zero at inference they form 902 tokens. Four encoder layers mix them, one decoder layer with 100 query slots reads the result, and a linear layer gives 100 joint targets, all of which the arm runs before it looks again.ACT turns one look into 100 actions.1 · observeCameras and jointsFRONTWRISTTOPj1j2j3j4j5gripz = 0STYLE LATENT3 × 480×640 RGB · state 62 · image backboneResNet18 per camera646412825651215 × 20 GRID512 × 15 × 20 → 300 tokens3 · encoder inputOne token sequencez​sfront 300wrist 300top 300902 TOKENS(2 + 300 V) × 5124 · transformer encoderMix every tokenself-attention→ feed-forward× 4each tokensees all 9024 layers · d 512 · 8 heads5 · decoder and action head100 slots, 100 actions100 QUERY SLOTScross-attention to tokens100 × 61 layer → Linear 512 → 66 · executeRun all 100, then look0 s3.33 s, then look again30 Hz · no re-planning in betweensource · lerobot/policies/act/modeling_act.py · click a block or use the arrow keys \ No newline at end of file diff --git a/_includes/robot-policy/act-training.svg b/_includes/robot-policy/act-training.svg index ef183e0a..75ca9892 100644 --- a/_includes/robot-policy/act-training.svg +++ b/_includes/robot-policy/act-training.svg @@ -1,306 +1 @@ - - ACT training step - A batch of recorded images, joint states and 100-step action chunks. A CVAE encoder reads the true actions and gives the mean and variance of a 32-dimensional style latent; z is sampled with the reparameterization trick and becomes the first token of the policy. The loss is the masked L1 error of the predicted chunk plus ten times the KL divergence; AdamW updates all weights at a constant learning rate. - - - - - - - ACT learns to rebuild the recorded chunk. - - - - - - - - - - - 1 · data - - - - A batch of demonstrations - - images · V × 3 × 480 × 640 - - joint state · 6 - - actions · 100 × 6, padded past the end - thanhkt/nexarm_stack_bowls · last 10 % held out - batch 2 × 8 accumulation = 16 - - - 2 · CVAE encoder · training only - - - - Encode the true chunk - - CLS - - s - - a - - a - - a - - a - - a - - a - - a - - pad - - - 4 encoder layers · d 512 - - - μ, log σ² · 32 each - [CLS] output → Linear 512 → 64 - - - 3 · sample z - - - - Sample a style - - - - prior N(0, 1) - q(z | s, a) - one of 32 dimensions · at inference z = 0 - z = μ + σ ⊙ ε, ε ~ N(0, I) - - - 4 · ACT network with z - - - - The policy, given z - - z - FIRST TOKEN - - - ResNet18 - - - encoder ×4 - - - decoder ×1 - - - - - - - - - - - - - - - - - - - - â · 100 × 6 - the network from the Inference tab - z replaces the zero latent - - - 5 · loss - - - - Rebuild the chunk - j1 - - - - - - - - - - - - - - - - - - - - - - - j2 - - - - - - - - - - - - - - - - - - - - - - - j3 - - - - - - - - - - - - - - - - - - - - - - - j4 - - - - - - - - - - - - - - - - - - - - - - - j5 - - - - - - - - - - - - - - - - - - - - - - - grip - - - - - - - - - - - - - - - - - - - - - - - darker = larger |â − a| · grey = padded - + KL(q ‖ N(0, I)) summed over 32 dims - L1 + 10 · KL - - - 6 · optimizer step - - - - Update every weight - - backward × 8 - - - clip 10 - - - AdamW - - - lr - 1e-5, constant - step 100k - weight decay 1e-4 · no schedule - - source · lerobot train, ACT preset · 100,000 steps · click a block or use the arrow keys - +ACT training stepA batch of recorded images, joint states and 100-step action chunks. A CVAE encoder reads the true actions and gives the mean and variance of a 32-dimensional style latent; z is sampled with the reparameterization trick and becomes the first token of the policy. The loss is the masked L1 error of the predicted chunk plus ten times the KL divergence; AdamW updates all weights at a constant learning rate.ACT learns to rebuild the recorded chunk.1 · dataA batch of demonstrationsimages · V × 3 × 480 × 640joint state · 6actions · 100 × 6, padded past the endthanhkt/nexarm_stack_bowls · last 10 % held outbatch 2 × 8 accumulation = 162 · CVAE encoder · training onlyEncode the true chunkCLSsaaaaaaapad4 encoder layers · d 512μ, log σ² · 32 each[CLS] output → Linear 512 → 643 · sample zSample a styleprior N(0, 1)q(z | s, a)one of 32 dimensions · at inference z = 0z = μ + σ ⊙ ε, ε ~ N(0, I)4 · ACT network with zThe policy, given zzFIRST TOKENResNet18encoder ×4decoder ×1â · 100 × 6the network from the Inference tabz replaces the zero latent5 · lossRebuild the chunkj1j2j3j4j5gripdarker = larger |â − a| · grey = padded+ KL(q ‖ N(0, I)) summed over 32 dimsL1 + 10 · KL6 · optimizer stepUpdate every weightbackward × 8clip 10AdamWlr1e-5, constantstep 100kweight decay 1e-4 · no schedulesource · lerobot train, ACT preset · 100,000 steps · click a block or use the arrow keys \ No newline at end of file diff --git a/_includes/robot-policy/bsp-pipeline.svg b/_includes/robot-policy/bsp-pipeline.svg index e95278d7..adbdce53 100644 --- a/_includes/robot-policy/bsp-pipeline.svg +++ b/_includes/robot-policy/bsp-pipeline.svg @@ -1,177 +1 @@ - - B-spline Policy inference on the NexArm - The cameras and the joint state condition a Diffusion Policy UNet, which denoises one cubic B-spline segment: 16 knots in one column and the control points in the others. Each basis function is non-zero over four knot spans, so moving one control point changes the curve only locally. The controller samples the curve at 100 Hz and can traverse it n times faster; the next segment is entered at the point that matches the last executed action. - - - - - - - BSP predicts a curve, not a list of actions. - - - - - - - - - - - 1 · observe - - - - Cameras and joints - - FRONT - - WRIST - - TOP - - j1 - - j2 - - j3 - - j4 - - j5 - - grip - policy 10 Hz - control 100 Hz - last 2 frames · joint state - - - 2 · Diffusion Policy - - - - Denoise a curve - - ResNet18 - - condition - - 1-D UNet - - - 16 × (1 + D) - - × 16 DDIM steps - knot + control point - noise → [U ; C] in 16 steps - - - 3 · B-spline parameters - - - - 16 knots, their control points - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 3 + 3 support · 12 control points - - - 4 · the curve is a weighted sum - - - - One point moves 4 spans - - - - - - - - - - - - - - - - - a(u) = Σ Ni,3​(u) ci​ - - - 5 · sample at 100 Hz - - - - Same path, n× faster - 1× - - - 94 samples · 0.93 s - 2× - - - 47 samples · 0.47 s - aexec​(t) = a(n t) - - - 6 · join the next segment - - - - Join where the arm is - - - - - - - window · t* = best match - t* = argmin MSE(Snew​(t), alast​) - - source · Han, Xiong et al. 2026, B-spline Policy (arXiv 2607.09648) · toy joint trajectory · click a block or use the arrow keys - +B-spline Policy inference on the NexArmThe cameras and the joint state condition a Diffusion Policy UNet, which denoises one cubic B-spline segment: 16 knots in one column and the control points in the others. Each basis function is non-zero over four knot spans, so moving one control point changes the curve only locally. The controller samples the curve at 100 Hz and can traverse it n times faster; the next segment is entered at the point that matches the last executed action.BSP predicts a curve, not a list of actions.1 · observeCameras and jointsFRONTWRISTTOPj1j2j3j4j5grippolicy 10 Hzcontrol 100 Hzlast 2 frames · joint state2 · Diffusion PolicyDenoise a curveResNet18condition1-D UNet16 × (1 + D)× 16 DDIM stepsknot + control pointnoise → [U ; C] in 16 steps3 · B-spline parameters16 knots, their control points3 + 3 support · 12 control points4 · the curve is a weighted sumOne point moves 4 spansa(u) = Σ Ni,3​(u) ci​5 · sample at 100 HzSame path, n× faster1×94 samples · 0.93 s2×47 samples · 0.47 saexec​(t) = a(n t)6 · join the next segmentJoin where the arm iswindow · t* = best matcht* = argmin MSE(Snew​(t), alast​)source · Han, Xiong et al. 2026, B-spline Policy (arXiv 2607.09648) · toy joint trajectory · click a block or use the arrow keys \ No newline at end of file diff --git a/_includes/robot-policy/bsp-training.svg b/_includes/robot-policy/bsp-training.svg index 78ecdfad..f847b4b2 100644 --- a/_includes/robot-policy/bsp-training.svg +++ b/_includes/robot-policy/bsp-training.svg @@ -1,325 +1 @@ - - B-spline Policy training - Each demonstration is fitted with a cubic B-spline by inserting knots where the error is largest until it is below a tolerance; knots cluster where the motion bends. At each observation the target is the next 16 knots and their control points, which cover a long time where the motion is simple and a short time where it is detailed. A Diffusion Policy UNet learns to denoise them. A looser tolerance compresses more with little change in score (Push-T, Table 3). - - - - - - - BSP learns the curve behind each demonstration. - - - - - - - - - - - 1 · demonstrations - - - - A list of joint targets - - - REACH - - GRASP - - CARRY - - PLACE - one joint · 50 Hz · 8 s · 401 samples - - - 2 · adaptive B-spline fit - - - - Knots go where it bends - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 12 · GRASP - 7 · CARRY - max error ≤ ε = 0.004 · 41 knots - - - 3 · fixed-size target - - - - The next 16 knots - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - OVER THE CARRY - 16 knots · 3.04 s - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - OVER THE PLACE - 16 knots · 2.03 s - same size, different time span - - - 4 · denoising loss - - - - Learn to remove noise - - [U ; C] - + - - ε - - - noisy, step k - - - UNet → εθ​ - CONDITIONED ON - IMAGES + JOINT STATE - loss = ‖εθ​ − ε‖² - DDIM · 100 train steps · ε-prediction - - - 5 · 16 actions vs 16 knots - - - - Detail where it is needed - - - - - - - - - - - - - - - - - - - - CHUNK · 16 × 0.1 s, ALWAYS EVEN - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - KNOTS · DENSE WHERE IT IS FINE - adaptive temporal resolution - - - 6 · fit tolerance ε - - - - Looser fit, same score - - - - 1.12× - ε 0.5 - 0.52 - - 1.34× - ε 1 - 0.65 - - 1.73× - ε 2 - 0.62 - - 2.36× - ε 4 - 0.63 - - 3.34× - ε 8 - 0.65 - SCORE - Push-T · paper Table 3 - - source · Han, Xiong et al. 2026, B-spline Policy (arXiv 2607.09648), Alg. 1, App. A.1, Table 3 · toy joint trajectory - +B-spline Policy trainingEach demonstration is fitted with a cubic B-spline by inserting knots where the error is largest until it is below a tolerance; knots cluster where the motion bends. At each observation the target is the next 16 knots and their control points, which cover a long time where the motion is simple and a short time where it is detailed. A Diffusion Policy UNet learns to denoise them. A looser tolerance compresses more with little change in score (Push-T, Table 3).BSP learns the curve behind each demonstration.1 · demonstrationsA list of joint targetsREACHGRASPCARRYPLACEone joint · 50 Hz · 8 s · 401 samples2 · adaptive B-spline fitKnots go where it bends12 · GRASP7 · CARRYmax error ≤ ε = 0.004 · 41 knots3 · fixed-size targetThe next 16 knotsOVER THE CARRY16 knots · 3.04 sOVER THE PLACE16 knots · 2.03 ssame size, different time span4 · denoising lossLearn to remove noise[U ; C]+εnoisy, step kUNet → εθ​CONDITIONED ONIMAGES + JOINT STATEloss = ‖εθ​ − ε‖²DDIM · 100 train steps · ε-prediction5 · 16 actions vs 16 knotsDetail where it is neededCHUNK · 16 × 0.1 s, ALWAYS EVENKNOTS · DENSE WHERE IT IS FINEadaptive temporal resolution6 · fit tolerance εLooser fit, same score1.12×ε 0.50.521.34×ε 10.651.73×ε 20.622.36×ε 40.633.34×ε 80.65SCOREPush-T · paper Table 3source · Han, Xiong et al. 2026, B-spline Policy (arXiv 2607.09648), Alg. 1, App. A.1, Table 3 · toy joint trajectory \ No newline at end of file diff --git a/_includes/robot-policy/turbovla-pipeline.svg b/_includes/robot-policy/turbovla-pipeline.svg index dedf2439..2495cd4e 100644 --- a/_includes/robot-policy/turbovla-pipeline.svg +++ b/_includes/robot-policy/turbovla-pipeline.svg @@ -1,409 +1 @@ - - TurboVLA inference pipeline on the NexArm - Camera frames, the task instruction and the joint state go in. DINOv3 turns each 224 by 224 view into 196 tokens, a frozen BERT encodes the instruction, and six bi-attention layers let image and text tokens attend to each other. The joint state becomes two tokens. A three-layer decoder with 16 queries predicts 16 actions; the arm runs the first 8 and then looks again. - - - - - - - TurboVLA reads the task, then plans 16 steps. - - - - - - - - - - - - 1 · observe - - - - Look and read - - FRONT - - WRIST - - TOP - - “stack the bowls” - - j1 - - j2 - - j3 - - j4 - - j5 - - grip - 224×224 · text · state 6 - - - 2 · vision encoder - - - - DINOv3 per camera - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 196 tokens - per view - 768 → 256 - 3 views → 588 tokens - - - 3 · language encoder - - - - BERT reads the task - - stack - - the - - bowls - - - BERT · frozen - - - - - - - - L tokens × 256 - - - 4 · vision ↔ language - - - - Image and text, mixed - - - - - - - - - - - - - - - - - - - - shared logits - × 6 - 6 bi-attention layers - - - 5 · state tokens - - - - State → 2 tokens - - j1 - - j2 - - j3 - - j4 - - j5 - - grip - - - MLP - - - s₁ - - s₂ - 6 → 2 × 256 - - - 6 · action decoder - - - - 16 queries → 16 actions - - - - - - - - - - - - - - - - - 16 QUERIES - - - 3 decoder layers - - tanh - â · 16 × 6 in [−1, 1] - - - 7 · execute - - - - Run 8, look again - - - - - - - - - - - - - - - - - - - RUN 8 - DROPPED - 8 × 33 ms ≈ 0.27 s - - source · TurboVLA/turbovla/models · click a block or use the arrow keys - +TurboVLA inference pipeline on the NexArmCamera frames, the task instruction and the joint state go in. DINOv3 turns each 224 by 224 view into 196 tokens, a frozen BERT encodes the instruction, and six bi-attention layers let image and text tokens attend to each other. The joint state becomes two tokens. A three-layer decoder with 16 queries predicts 16 actions; the arm runs the first 8 and then looks again.TurboVLA reads the task, then plans 16 steps.1 · observeLook and readFRONTWRISTTOP“stack the bowls”j1j2j3j4j5grip224×224 · text · state 62 · vision encoderDINOv3 per camera196 tokensper view768 → 2563 views → 588 tokens3 · language encoderBERT reads the taskstackthebowlsBERT · frozenL tokens × 2564 · vision ↔ languageImage and text, mixedshared logits× 66 bi-attention layers5 · state tokensState → 2 tokensj1j2j3j4j5gripMLPs₁s₂6 → 2 × 2566 · action decoder16 queries → 16 actions16 QUERIES3 decoder layerstanhâ · 16 × 6 in [−1, 1]7 · executeRun 8, look againRUN 8DROPPED8 × 33 ms ≈ 0.27 ssource · TurboVLA/turbovla/models · click a block or use the arrow keys \ No newline at end of file diff --git a/_includes/robot-policy/turbovla-training.svg b/_includes/robot-policy/turbovla-training.svg index b80c007e..422819c7 100644 --- a/_includes/robot-policy/turbovla-training.svg +++ b/_includes/robot-policy/turbovla-training.svg @@ -1,253 +1 @@ - - TurboVLA training step - Samples are drawn from simulated and real recordings at a set ratio, images are resized to 224, state and actions min-max scaled to [-1, 1]. DINOv3, the interaction layers and the decoder are trained while BERT stays frozen. The loss is a masked L1 over 16 steps with an optional gripper weight; AdamW with warm-up and a cosine schedule updates the weights, and an exponential moving average of them is saved for rollout. - - - - - - - TurboVLA learns from mixed real and simulated runs. - - - - - - - - - - - 1 · data - - - - Real and sim, mixed - - simulation - - real - MuJoCo, randomized - teleoperated - - - - real with p = 0.5 - DomainMixSampler · batch 16 - - - 2 · collate and normalize - - - - Collate and normalize - - images → 224 × 224 - - state, actions → [−1, 1] - - instruction, kept as text - - - - - - - - - - - - - - - - - - - 16-step chunk · 3 padded - min / max over sim + real - - - 3 · forward pass - - - - Forward pass - - DINOv3 - - BERT · frozen - - - - interaction ×6 - - - decoder + MLP - bf16 autocast · BERT gets no grad - - - 4 · masked L1 loss - - - - Masked L1 - j1 - - - - - - - - - - - - - - - - - j2 - - - - - - - - - - - - - - - - - j3 - - - - - - - - - - - - - - - - - j4 - - - - - - - - - - - - - - - - - j5 - - - - - - - - - - - - - - - - - grip - - - - - - - - - - - - - - - - - bold row: gripper, optional weight w - Σ mask · w · |â − a| / valid - - - 5 · optimizer step - - - - Warm up, then decay - - backward - - - clip 1.0 - - - AdamW - - - lr - warm-up 1k - cosine → 5 % · 50k - AdamW · lr 5e-5 · LambdaLR - - - 6 · EMA and checkpoint - - - - Keep a smoothed copy - θema​ ← 0.999 θema​ + 0.001 θ - after every optimizer step - - {step}_ema_model.safetensors - - {step}_model.pt - - config.json - - stats_turbovla.json - every 5,000 steps · rollout uses EMA - - source · examples/nexarm/train_turbovla.py · 50,000 steps · click a block or use the arrow keys - +TurboVLA training stepSamples are drawn from simulated and real recordings at a set ratio, images are resized to 224, state and actions min-max scaled to [-1, 1]. DINOv3, the interaction layers and the decoder are trained while BERT stays frozen. The loss is a masked L1 over 16 steps with an optional gripper weight; AdamW with warm-up and a cosine schedule updates the weights, and an exponential moving average of them is saved for rollout.TurboVLA learns from mixed real and simulated runs.1 · dataReal and sim, mixedsimulationrealMuJoCo, randomizedteleoperatedreal with p = 0.5DomainMixSampler · batch 162 · collate and normalizeCollate and normalizeimages → 224 × 224state, actions → [−1, 1]instruction, kept as text16-step chunk · 3 paddedmin / max over sim + real3 · forward passForward passDINOv3BERT · frozeninteraction ×6decoder + MLPbf16 autocast · BERT gets no grad4 · masked L1 lossMasked L1j1j2j3j4j5gripbold row: gripper, optional weight wΣ mask · w · |â − a| / valid5 · optimizer stepWarm up, then decaybackwardclip 1.0AdamWlrwarm-up 1kcosine → 5 % · 50kAdamW · lr 5e-5 · LambdaLR6 · EMA and checkpointKeep a smoothed copyθema​ ← 0.999 θema​ + 0.001 θafter every optimizer step{step}_ema_model.safetensors{step}_model.ptconfig.jsonstats_turbovla.jsonevery 5,000 steps · rollout uses EMAsource · examples/nexarm/train_turbovla.py · 50,000 steps · click a block or use the arrow keys \ No newline at end of file diff --git a/assets/img/projects/6dof-robot/gantry-rig-cad-render.png b/assets/img/projects/6dof-robot/gantry-rig-cad-render.png index 428fd3fe..acaad7e5 100644 Binary files a/assets/img/projects/6dof-robot/gantry-rig-cad-render.png and b/assets/img/projects/6dof-robot/gantry-rig-cad-render.png differ diff --git a/assets/img/projects/aikc/homepage.png b/assets/img/projects/aikc/homepage.png index 9b3378c4..2b408c14 100644 Binary files a/assets/img/projects/aikc/homepage.png and b/assets/img/projects/aikc/homepage.png differ diff --git a/assets/img/projects/infi/analysis-data-points.png b/assets/img/projects/infi/analysis-data-points.png index e1dbe80c..2233694b 100644 Binary files a/assets/img/projects/infi/analysis-data-points.png and b/assets/img/projects/infi/analysis-data-points.png differ diff --git a/assets/img/projects/infi/analysis-final-stance.png b/assets/img/projects/infi/analysis-final-stance.png index c10ce644..e55e6c1b 100644 Binary files a/assets/img/projects/infi/analysis-final-stance.png and b/assets/img/projects/infi/analysis-final-stance.png differ diff --git a/assets/img/projects/infi/analysis-projection.png b/assets/img/projects/infi/analysis-projection.png index 821c6918..116a57d6 100644 Binary files a/assets/img/projects/infi/analysis-projection.png and b/assets/img/projects/infi/analysis-projection.png differ diff --git a/assets/img/projects/infi/analysis-scenario-matrix.png b/assets/img/projects/infi/analysis-scenario-matrix.png index cbe0c485..b0927267 100644 Binary files a/assets/img/projects/infi/analysis-scenario-matrix.png and b/assets/img/projects/infi/analysis-scenario-matrix.png differ diff --git a/assets/img/projects/infi/analysis-sources.png b/assets/img/projects/infi/analysis-sources.png index b80f69cb..e61e8598 100644 Binary files a/assets/img/projects/infi/analysis-sources.png and b/assets/img/projects/infi/analysis-sources.png differ diff --git a/assets/img/projects/infi/analysis-thesis.png b/assets/img/projects/infi/analysis-thesis.png index 25a638fd..2009d1c6 100644 Binary files a/assets/img/projects/infi/analysis-thesis.png and b/assets/img/projects/infi/analysis-thesis.png differ diff --git a/assets/img/projects/infi/landing.png b/assets/img/projects/infi/landing.png index 5ca0205e..94dfedac 100644 Binary files a/assets/img/projects/infi/landing.png and b/assets/img/projects/infi/landing.png differ diff --git a/assets/img/projects/infi/new-analysis.png b/assets/img/projects/infi/new-analysis.png index c14ef178..e756516f 100644 Binary files a/assets/img/projects/infi/new-analysis.png and b/assets/img/projects/infi/new-analysis.png differ diff --git a/assets/img/projects/infi/portfolio-rebalancing-scenarios.png b/assets/img/projects/infi/portfolio-rebalancing-scenarios.png index ccb2b93a..d6686c06 100644 Binary files a/assets/img/projects/infi/portfolio-rebalancing-scenarios.png and b/assets/img/projects/infi/portfolio-rebalancing-scenarios.png differ diff --git a/assets/img/projects/infi/portfolio-risk.png b/assets/img/projects/infi/portfolio-risk.png index ad7e35d7..b40482d6 100644 Binary files a/assets/img/projects/infi/portfolio-risk.png and b/assets/img/projects/infi/portfolio-risk.png differ diff --git a/assets/img/projects/stemfun/landing.png b/assets/img/projects/stemfun/landing.png index 7094e6d1..c874986a 100644 Binary files a/assets/img/projects/stemfun/landing.png and b/assets/img/projects/stemfun/landing.png differ diff --git a/assets/img/publication_preview/prism-animation.svg b/assets/img/publication_preview/prism-animation.svg index dc6f3f78..acd43076 100644 --- a/assets/img/publication_preview/prism-animation.svg +++ b/assets/img/publication_preview/prism-animation.svg @@ -1,321 +1 @@ - - - - - - - - - - - - - - - - - - 1 · DEPTH OF ANALYSIS - - - Classify ADUs - Claim vs Premise · Aspect Topic - - - - - Assess Grounding Level g(p) - - - L0: Vague - - - - L1: Internal - - - - L2: External - - - - - - Harmonic Mean Score - DoA = 2·(R_prem · S_depth) / (R_prem + S_depth) - - - - - - - - - 2 · NOVELTY ASSESSMENT - - - Extract Novelty Claims - Task · Contribution Anchors · Key Terms - - - - - Semantic Scholar Retrieval - MMR Diversified Pool B = {b₁...bₖ} - - - - - LLM Judge Verification - s(cᵢ,bⱼ) ∈ {-2, -1, 0, +1, +2} - - - - - - - - - -2 - -1 - 0 - +1 - +2 - - - - - - - - DATA SOURCE - 1,000 Papers · 200 per venue - - - ICLR 24/25/26 - - ICML 2025 - - NeurIPS 2025 - - SEA · TreeReview · CycleReviewer · DeepReview · Reviewer2 · Human - - - - - - - - - - - - - - - MANUSCRIPT - - - - - Manuscript - - - + - - - - - PEER REVIEW - - - - - Peer Review - - - - - - - - - DATA SEGMENTATION - Summary · Strengths · Weaknesses · Questions - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 3 · FLAW IDENTIFICATION & PRIORITY - - - Extract Flaw Arguments - Human + LLM Reviews → Parallel Parse - - - - - Consensus Verification - LLM Judge → Valid vs Hallucinated - - - Critical - - - - Minor - - - - - - Positional Recovery → Rank - Critical flaws should appear first - - - - - - - - #1 Critical - - - - - - - #2 Critical - - - - - - - #3 Minor - - - - - - - - - - 4 · MULTI-DIM CONSTRUCTIVENESS - - - Decompose into ARCs - Atomic Review Comments - - - - - Rate 5 Dimensions (0-2) - - D1 Actionability - - - - - - - D2 Specificity - - - - - - - D3 Justification - - - - - - - D4 Solution - - - - - - - D5 Tone - - - - - - - - - - MCS = (1/n) Σ CLC(cⱼ) ∈ [0,1] - - - - - - +1 · DEPTH OF ANALYSISClassify ADUsClaim vs Premise · Aspect TopicAssess Grounding Level g(p)L0: VagueL1: InternalL2: ExternalHarmonic Mean ScoreDoA = 2·(R_prem · S_depth) / (R_prem + S_depth)2 · NOVELTY ASSESSMENTExtract Novelty ClaimsTask · Contribution Anchors · Key TermsSemantic Scholar RetrievalMMR Diversified Pool B = {b₁...bₖ}LLM Judge Verifications(cᵢ,bⱼ) ∈ {-2, -1, 0, +1, +2}-2-10+1+2DATA SOURCE1,000 Papers · 200 per venueICLR 24/25/26ICML 2025NeurIPS 2025SEA · TreeReview · CycleReviewer · DeepReview · Reviewer2 · HumanMANUSCRIPTManuscript+PEER REVIEWPeer ReviewDATA SEGMENTATIONSummary · Strengths · Weaknesses · Questions3 · FLAW IDENTIFICATION & PRIORITYExtract Flaw ArgumentsHuman + LLM Reviews → Parallel ParseConsensus VerificationLLM Judge → Valid vs HallucinatedCriticalMinorPositional Recovery → RankCritical flaws should appear first#1 Critical#2 Critical#3 Minor4 · MULTI-DIM CONSTRUCTIVENESSDecompose into ARCsAtomic Review CommentsRate 5 Dimensions (0-2)D1 ActionabilityD2 SpecificityD3 JustificationD4 SolutionD5 ToneMCS = (1/n) Σ CLC(cⱼ) ∈ [0,1] \ No newline at end of file diff --git a/lighthouse_results/desktop/pagespeed.svg b/lighthouse_results/desktop/pagespeed.svg index 0683494c..9fd0edfe 100644 --- a/lighthouse_results/desktop/pagespeed.svg +++ b/lighthouse_results/desktop/pagespeed.svg @@ -1 +1 @@ -95Performance100Accessibility100Best Practices100SEONAPWA0-4950-8990-100 +95Performance100Accessibility100Best Practices100SEONAPWA0-4950-8990-100 \ No newline at end of file diff --git a/lighthouse_results/mobile/pagespeed.svg b/lighthouse_results/mobile/pagespeed.svg index d14bbec7..5c1a4e84 100644 --- a/lighthouse_results/mobile/pagespeed.svg +++ b/lighthouse_results/mobile/pagespeed.svg @@ -1 +1 @@ -92Performance100Accessibility100Best Practices100SEONAPWA0-4950-8990-100 +92Performance100Accessibility100Best Practices100SEONAPWA0-4950-8990-100 \ No newline at end of file diff --git a/readme_preview/blog.png b/readme_preview/blog.png index 0f368285..2eb71611 100644 Binary files a/readme_preview/blog.png and b/readme_preview/blog.png differ diff --git a/readme_preview/code.png b/readme_preview/code.png index 6dd6f84d..14f22e99 100644 Binary files a/readme_preview/code.png and b/readme_preview/code.png differ diff --git a/readme_preview/cv.png b/readme_preview/cv.png index b92c9f08..29dd4240 100644 Binary files a/readme_preview/cv.png and b/readme_preview/cv.png differ diff --git a/readme_preview/dark.png b/readme_preview/dark.png index 311479ce..66df1157 100644 Binary files a/readme_preview/dark.png and b/readme_preview/dark.png differ diff --git a/readme_preview/distill.png b/readme_preview/distill.png index 99d7847f..7a5bab37 100644 Binary files a/readme_preview/distill.png and b/readme_preview/distill.png differ diff --git a/readme_preview/jupyter.png b/readme_preview/jupyter.png index 88909187..d1e3ee99 100644 Binary files a/readme_preview/jupyter.png and b/readme_preview/jupyter.png differ diff --git a/readme_preview/light.png b/readme_preview/light.png index 985686a3..db96fc5a 100644 Binary files a/readme_preview/light.png and b/readme_preview/light.png differ diff --git a/readme_preview/math.png b/readme_preview/math.png index 7f41e3ee..6f45df27 100644 Binary files a/readme_preview/math.png and b/readme_preview/math.png differ diff --git a/readme_preview/people.png b/readme_preview/people.png index 3e3687c2..8ed3f247 100644 Binary files a/readme_preview/people.png and b/readme_preview/people.png differ diff --git a/readme_preview/photos-screenshot.png b/readme_preview/photos-screenshot.png index 3dce0a12..adad2fd8 100644 Binary files a/readme_preview/photos-screenshot.png and b/readme_preview/photos-screenshot.png differ diff --git a/readme_preview/projects.png b/readme_preview/projects.png index 375a8981..7f647d82 100644 Binary files a/readme_preview/projects.png and b/readme_preview/projects.png differ diff --git a/readme_preview/publications.png b/readme_preview/publications.png index f1d0068e..bf7eb578 100644 Binary files a/readme_preview/publications.png and b/readme_preview/publications.png differ diff --git a/readme_preview/repositories.png b/readme_preview/repositories.png index e1a7ac95..37f1a6c2 100644 Binary files a/readme_preview/repositories.png and b/readme_preview/repositories.png differ