All Posts
All notes and posts, grouped only by date order
- Learning Action Priors for Cross-embodiment Robot Manipulation
Pretrained VLM에 아직 motor structure를 배우지 못한 action head를 바로 붙여 joint train하는 대신, state-action trajectory만으로 flow-matching action encoder-decoder를 먼저 pretrain한 뒤 decoder initialization, decaying latent distillation, history compression을 통해 VLA에 이식하는 cross-embodiment policy training framework
Koreansuccess-rateVLAcross-embodimentfine-tuningcomponent-scratch-training - Grounding Generative Policies in Physics: Optimization-Guided Diffusion for Robot Control
Frozen task-space diffusion policy의 DDIM sampling noise를 무작위로 뽑는 대신, robot reachability·collision·controller trackability를 만족하도록 최적화하여 cross-embodiment deployment를 수행하는 inference-time constrained diffusion method
Koreaninference-timetraining-freediffusion-policycross-embodiment - InSight: Self-Guided Skill Acquisition via Steerable VLAs
기존 demonstration을 자동으로 primitive 단위로 분해해 pretrained π0.5를 primitive-steerable policy로 만들고, novel task에서 VLM이 발견한 missing primitive를 single-axis controller로 자율 수집·검증한 뒤 VLA에 재학습하여 영속적인 skill vocabulary로 편입하는 VLM-guided continual skill acquisition framework
Koreansuccess-rateVLAfine-tuningauxiliary-module-training - PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models
pretrained VLA를 두 단계의 GRPO 기반 RL post-training으로 fine-tuning하여, 한 번의 inference에서 안전하게 실행할 수 있는 action chunk 길이를 늘리고 전체 physical control step은 줄임
Koreansuccess-rateVLAfine-tuning - SPACE: Enabling Learning from Cross-Robot Data Toward Generalist Policies
VLA가 robot-specific control command 대신 실제로 달성해야 할 6-DoF Cartesian end-effector displacement를 예측하게 하고, target robot마다 선형 Action Adapter를 offline calibration과 online LMS로 적응시켜 cross-embodiment·cross-hardware·deployment dynamics shift에 강한 execution interface를 만든다
Koreansuccess-rateVLAfine-tuningauxiliary-module-trainingcross-embodiment - World Value Models for Robotic Manipulation
Pretrained Wan2.2 video world model을 robot video로 jointly fine-tune하면서 별도의 lightweight value DiT를 Mixture-of-Transformers로 결합해, video와 language로부터 4-frame task-progress chunk를 flow matching으로 생성하고 그 progress 변화량으로 suboptimal data를 filtering·reweighting하는 generalist robotic value model
Koreansuccess-ratefoundation-modelfine-tuning - FlowDPG: Deterministic Policy Gradient on Flow Matching Policies for Real-World Manipulation
flow matching robot policy의 중간 noisy action을 clean action chunk로 한 번에 projection한 뒤, 그 지점의 critic gradient를 value-improved velocity target으로 distillation하여 전체 denoising ODE를 backpropagation하지 않고도 offline-to-online real-world RL을 수행
Koreansuccess-ratefine-tuningauxiliary-module-trainingcomponent-scratch-training - UniFS: Unified Fast-to-Slow Hierarchical Architecture for Vision-Language-Action Models
pretrained VLM과 action expert의 각 layer group을 서로 다른 주기로 실행·cache하도록 학습하고, VLM feature와 action decoding stage의 연결 순서를 뒤집어, VLA-Adapter의 success rate를 높이면서 평균 inference latency를 줄인 scheduler-aware VLA architecture
Koreansuccess-rateinference-timeVLAfine-tuningcomponent-scratch-trainingscheduler-training - UniviewVLA: A Unified Multiview Vision-Language-Action Model with World Modeling
agent-view와 wrist-view의 두 프레임만으로 candidate auxiliary-views의 다음 장면 token을 생성하고, motion-relevant token 16개로 압축한 뒤 action entropy가 가장 낮은 view를 선택해 FAST action token을 생성하는 autoregressive multiview VLA
Koreansuccess-rateVLAfine-tuning - ENPIRE: Agentic Robot Policy Self-Improvement in the Real World
coding agent가 실제 로봇의 reset → rollout → verification → policy/code refinement research loop를 직접 운영하고, 여러 robot–agent worker가 Git으로 실험 지식을 공유하면서 task policy를 자동 개선하게 만든 physical autoresearch harness
Koreansuccess-ratetraining-dataVLA - Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think
pretrained π0·GR00T-N1.5·SmolVLA의 VLM backbone과 continuous-action head에서 Centered Kernel Alignment로 표현이 거의 변하지 않는 연속 Transformer layer를 찾아 fine-tuning 전에 정적으로 제거하고, 남은 작은 모델을 downstream fine-tuning하여 학습·추론 비용을 함께 줄이는 VLA structural pruning method
Koreaninference-timeVLAfine-tuning - ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?
pretrained image-editing model을 robot policy backbone으로 fine-tuning하고, future video를 생성하는 대신 single future endpoint를 학습할 때 형성되는 layer-wise KV cache를 flow-matching Action Expert에 전달하여 action chunk를 생성하는 경량 WAM
Koreaninference-timeWAMfine-tuning - Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection
frozen flow-matching robot policy의 initial action noise를 backward ODE inversion과 repainting으로 조정하여, 이미 실행된 action prefix와 새 action chunk를 gradient·retraining 없이 연속적으로 연결하는 asynchronous inference method
Koreaninference-timesuccess-rateVLAtraining-free - 5. Valid Image Embedding Batching
observation image를 각각 embed하지 말고 한번에 embedding을 구해서 나중에 split
KoreanPythonProfiling - Weekly Review #3
2026.06.15 ~ 2026.06.19
Korean - SC3-Eval: Evaluating Robot Foundation Models via Self-Consistent Video Generation
pretrained Cosmos3-Nano video foundation model을 forward dynamics, inverse dynamics, cross-view inpainting의 세 mode로 공동 fine-tuning하고, inference에서는 commanded action과 generated video에서 inverse dynamics로 복원한 action의 불일치를 rollout reliability signal로 사용해 frozen VLA policy를 multi-view video world model 안에서 closed-loop 평가하는 method
Koreansuccess-rateVLAWAMbenchmarkfine-tuningauxiliary-module-training - Do as I Do: Dexterous Manipulation Data from Everyday Human Videos
monocular RGB human manipulation video를 4D hand–object trajectory로 복원하고, pretrained SAM 3D를 training-free guided flow sampling으로 object tracker처럼 재활용한 뒤, MuJoCo Warp의 dynamics-aware sampling optimization으로 22-DoF Sharpa Wave hand가 실행할 수 있는 robot trajectory로 변환하는 offline robot-data engine
Koreansuccess-ratetraining-datatraining-freecross-embodiment - DREAM-Chunk: Reactive Action Chunking with Latent World Model
frozen action-chunking VLA가 샘플링한 N개 candidate chunk의 latent future를 lightweight world model로 예측하고, 매 control step마다 현재 observation과 가장 가까운 phase-aligned dreamed state의 action으로 전환해 VLA를 다시 호출하지 않고 within-chunk reactivity를 높이는 test-time scaling method
Koreaninference-timesuccess-rateVLAauxiliary-module-training - MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction
RGB history, object 위의 2D query points와 corresponding initial 3D coordinates, language instruction을 입력받아 object-attached point들의 미래 3D world-frame trajectory를 예측하도록 Molmo2를 대규모 human/robot/in-the-wild video로 pretrain하고, 이 motion prior가 robot policy initialization과 video generation guidance로 전이됨을 보임
Koreansuccess-ratefoundation-modeltraining-datacross-embodiment - Object-Centric Residual RL for Zero-Shot Sim-to-Real VLA Enhancement
Real-robot demonstrations로 fine-tune한 VLA를 고정한 뒤, task-relevant object 6-DoF pose·proprioception·현재 base VLA action만 입력받는 lightweight residual RL policy를 simulation에서 학습하고 real robot에 adaptation 없이 결합해, FR3 5-task 평균 real success rate를 42%에서 76%로 높인 sim-to-real VLA enhancement framework
Koreansuccess-rateVLAauxiliary-module-trainingfine-tuningtraining-datasim2real - PAIWorld: A 3D-Consistent World Foundation Model for Robotic Manipulation
pretrained 14B flow-matching video DiT에 Geometry-Aware Cross-View Attention, camera-aware Geo-RoPE, Depth Anything 3 기반 Latent 3D-REPA를 결합해 여러 로봇 카메라의 미래 영상을 3D-consistent하게 생성하고, action-conditioned rollout을 WAM의 world-prediction backbone으로 활용할 수 있는 multi-view world foundation model
Koreansuccess-ratefoundation-modelWAM - ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining
대규모 egocentric human video를 robot-compatible pseudo-action으로 변환하고, camera-space action / morphology conditioning / time-aligned chunking / reliability-aware auxiliary loss를 결합해 human + robot + simulation 데이터를 함께 VLA pretraining에 쓰는 unified VLA pretraining framework
Koreansuccess-rateVLAfine-tuningtraining-datacross-embodiment - LAGO Policy: Latency-Aware Asynchronous Diffusion Policies with Goal-Directed Collision-Free Planning for Smooth Manipulation
asynchronous inference로 실행되는 Diffusion Policy의 chunk boundary jerk와 obstacle collision 문제를 latency-aware classifier-free guidance, demonstration-derived goal prediction, collision-free trajectory optimization, spatial-temporal smoothing으로 줄이는 real-robot manipulation policy
Koreaninference-timesuccess-ratediffusion-policyauxiliary-module-trainingscratch-training - Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models
Qwen-VL 기반 VLA에 canonical state/action alignment, camera-frame EEF action, in-context policy adaptation, Human-to-Robot synthesis를 결합해 heterogeneous robot manipulation data를 coherent하게 scale하고 OOD task/scene·instruction·cross-embodiment generalization을 끌어올린 robot manipulation foundation model
Koreansuccess-ratefoundation-modelVLAtraining-datacross-embodiment - Uncertainty Quantification for Flow-Based Vision-Language-Action Models
flow matching 기반 VLA의 action generation ODE에서 ensemble velocity field disagreement(VFD)를 측정해 epistemic uncertainty를 추정하고, 이를 failure detection과 SAVE active fine-tuning data acquisition에 사용해 expert demonstration sample efficiency를 높임
Koreaninference-timeVLAfine-tuningtraining-data - WAM-RL: World-Action Model Reinforcement Learning with Reconstruction Rewards and Online Video SFT
pretrained WAM에서 actor만 RL fine-tuning하지 않고, successful online rollout으로 world model을 KL-regularized video SFT하며, actor는 imagined future와 executed future의 reconstruction consistency reward로 RL update하는 WAM post-training framework
Koreansuccess-rateWAMfine-tuning - Where Should Action Generation Begin? A Learnable Source Prior for Generative Robot Policies
flow matching 기반 generative robot policy의 action generation source를 observation-independent Gaussian noise에서 proprioception-conditioned learnable Gaussian prior로 바꾸고, 같은 source prior가 diffusion-bridge generator에도 plug-in될 수 있음을 보인 source-prior learning method
Koreansuccess-ratescratch-trainingdiffusion-policy - Visual Verification Enables Inference-time Steering and Autonomous Policy Improvement
pretrained generalist robot policy를 stochastic action generator로 사용하고, geometric verifier로 Best-of-N action chunk를 선택한 뒤, 성공한 verified rollout을 BC fine-tuning data로 재사용하는 inference-time steering + autonomous policy improvement framework
Koreaninference-timesuccess-rateVLAfine-tuningWriting - 4. Prefix Fixed-Cost Breakdown and Masked Image Skip
prefix fixed cost를 image embedding과 prefill 단계로 분해한 뒤, LIBERO에서 mask 처리된 right-wrist image branch가 여전히 vision tower를 통과하는 낭비를 찾아 제거
KoreanPythonProfiling - Acting While Understanding: Asynchronous Semantic-Action Decoupling for Real-Time Vision-Language-Action Models
VLA 내부 semantic-action interface를 slow semantic understanding과 fast action generation으로 분리하고, stale semantic cache를 action history와 delay-aware training으로 보완해 full VLA를 control rate로 돌리지 않는 high-frequency state-feedback VLA deployment framework
Koreaninference-timeVLAfine-tuningcomponent-scratch-training - Geometric Action Model for Robot Policy Learning
pretrained Geometric Foundation Model(GFM)을 단순 feature extractor가 아니라 robot policy backbone 자체로 재활용해, GFM latent space 안에서 future geometry와 action chunk를 함께 예측하는 geometry-grounded World-Action Policy
Koreansuccess-rateVLAWAMfoundation-modelfine-tuningcomponent-scratch-training - Inference-time Policy Steering via Vision and Touch
frozen diffusion robot policy의 weights는 바꾸지 않고, action-conditioned visuo-tactile latent world model로 후보 action chunk의 future outcome을 예측한 뒤, long-horizon vision으로 global action mode를 선택하고 short-horizon touch로 local contact execution을 diffusion editing하는 inference-time steering method
Koreaninference-timesuccess-rateWAMdiffusion-policyauxiliary-module-training - Retrieve, Don’t Retrain: Extending Vision-Language-Action Models to New Tasks at Test Time
VLA/WAM policy를 새 task마다 다시 fine-tuning하지 않고, 저비용 pool embodiment demonstration을 retrieval pool에 추가한 뒤 frozen policy가 매 control step마다 retrieved trajectory를 조건으로 action chunk를 생성하게 만든 test-time task adaptation method
Koreaninference-timesuccess-rateVLAWAMfine-tuningtraining-freecross-embodiment - T-Rex: Tactile-Reactive Dexterous Manipulation
tactile-free human egocentric pretraining으로 얻은 visuomotor prior를 tactile-rich robot mid-training으로 contact dynamics에 맞춘 뒤, slow action expert와 fast tactile expert를 cascaded flow matching으로 연결해 action chunk 내부에서도 tactile feedback에 반응하는 tactile-reactive dexterous VLA
Koreansuccess-rateVLAfoundation-modelfine-tuningauxiliary-module-trainingtraining-dataMoE - 2. Limited Closed-loop Reproduction, Route-level Profiling, and Wrist-camera Robustness
A limited closed-loop reproduction and probing project for Realtime-VLA FLASH on Runpod L40S: official checkpoint conversion, LIBERO Goal baseline, synchronized route-level profiling, wrist-camera dropout robustness, and a minimal WristHealthGuard extension
EnglishProfilingPython - Elastic Queries Reinforcement Learning: Self-Aware Policy Execution for VLA Models
frozen flow-based VLA는 그대로 둔 채, lightweight RL adaptor가 매 query마다 latent steering w, denoising steps K, execution chunk length C를 동적으로 선택해 hard state에서는 더 많은 compute와 잦은 replanning을, easy state에서는 낮은 compute와 긴 open-loop execution을 수행하도록 만드는 elastic VLA execution framework
Koreaninference-timesuccess-rateVLAscheduler-trainingauxiliary-module-training - ReactVLA: Fast and Lightweight Reactive Robot Manipulation via Improved Mean Flow Action Generation
diffusion / flow 기반 VLA policy의 inference latency 병목을 줄이기 위해, action generation을 improved Mean Flow(iMF) 기반 one-to-few-step continuous action chunk generation으로 바꾸고 Attention Residuals(AttnRes) Transformer를 결합한 low-latency reactive robot manipulation policy
Koreaninference-timeVLAcomponent-scratch-training - WAM4D: Fast 4D World Action Model via Spatial Register Tokens
4D geometry를 inference-time output으로 직접 만들지 않고, training-time spatial register token으로 future depth를 예측하게 만들어 geometric foundation prior를 causal video-action WAM에 distill한 뒤, deploy 시 geometry branch를 제거해 action chunk를 빠르게 생성
Koreansuccess-rateWAMfine-tuningauxiliary-module-trainingcomponent-scratch-training - µ0: A Scalable 3D Interaction-Trace World Model
pretraining 단계에서는 action-labeled robot data 없이 heterogeneous videos에서 추출한 semantic 3D interaction traces를 학습하고, downstream에서는 frozen trace world model의 hidden features를 action expert에 주입해 robot policy를 만드는 3D trace-space world model
Koreansuccess-rateWAMfoundation-modeltraining-datacomponent-scratch-training - 1. Runpod Server Manifest
public-safe Runpod GPU server snapshot after initial setup
English - Weekly Review #2
2026.06.08 ~ 2026.06.12
Korean - EgoEngine: From Egocentric Human Videos to High-Fidelity Dexterous Robot Demonstrations
egocentric human manipulation video를 digital twin 기반으로 변환해, robot observation video와 실행 가능한 로봇 action trajectory를 함께 생성하고, 이를 이용해 real-robot dexterous visuomotor policy를 학습하는 human-video-to-robot-demo data engine
Koreansuccess-ratetraining-dataauxiliary-module-trainingcross-embodiment - Improving Robotic Generalist Policies via Flow Reversal Steering
coarse semantic action을 frozen flow-matching VLA의 역방향 ODE로 latent noise에 매핑한 뒤 다시 denoise해, generalist policy prior 안의 더 정교한 action mode를 호출하는 training-free steering 방법
Koreansuccess-rateinference-timeVLAauxiliary-module-trainingtraining-free - WEAVER, Better, Faster, Longer: An Effective World Model for Robotic Manipulation
multi-view RGB + proprioception + action chunk를 입력으로 미래 latent rollout과 reward/value를 빠르게 예측해, π0.5 같은 VLA policy의 offline evaluation, synthetic-data policy improvement, test-time best-of-N planning을 가능하게 만든 action-conditioned latent world model
Koreaninference-timesuccess-rateWAMVLAfine-tuningauxiliary-module-trainingtraining-data - Ambient Diffusion Policy: Imitation Learning from Suboptimal Data in Robotics
suboptimal / OOD robot demonstrations를 Diffusion Policy 학습에 그냥 섞지 않고, diffusion timestep에 따라 “쓸 수 있는 구간”을 제한해 유용한 global plan 또는 local motion primitive만 뽑아 쓰는 imitation learning 방법
Koreansuccess-ratediffusion-policyscratch-trainingtraining-data - Dynamic Execution Horizon Prediction for Chunk-based Robot Policies
pretrained action-chunking robot policy의 action generator는 완전히 고정하고, 현재 observation과 예측된 action chunk를 보고 “이번에 몇 step을 open-loop로 실행할지”를 PPO로 학습하는 lightweight execution-horizon predictor
Koreaninference-timesuccess-ratediffusion-policyscheduler-trainingauxiliary-module-training - DAM-VLA: Decoupled Asynchronous Multimodal Vision Language Action model
VLA의 synchronous clock 가정이 contact-rich manipulation의 multi-rate sensor structure와 맞지 않는다고 보고, modality별 asynchronous latent buffer + gated cross-attention으로 X-VLA를 100 Hz controller 기반 closed-loop execution에 맞춘다
KoreanVLAfine-tuningcomponent-scratch-training - Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination
WAM의 미래 영상 예측을 photorealistic video generation이 아니라 action generation을 돕는 저비용 coarse future guidance로 재정의하고, compact video expert + low-resolution future latent + asymmetric video-action denoising으로 약 1B 규모에서 real-world policy inference latency를 약 98 ms/chunk까지 낮춤
Koreaninference-timesuccess-rateWAMfine-tuningcomponent-scratch-training - SARM2: Multi-Task Stage Aware Reward Modeling for Self Improving Robotic Manipulation
long-horizon robotic manipulation에서 VLA policy의 self-improvement를 위해, action-primitive stage estimator와 multi-gate MoE value head로 dense reward/value model을 만들고, 이를 SPIRAL의 offline-to-online residual RL data flywheel에 통합한다
Koreansuccess-rateVLAfine-tuningauxiliary-module-trainingMoE - AHA-WAM: Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing
Video-DiT world planner는 low-frequency로 long-horizon latent context를 만들고, Action-DiT executor는 OVCR로 최신 observation에 맞게 context를 보정해 short action chunk를 high-frequency closed-loop로 실행하는 asynchronous WAM
KoreanWAMinference-timesuccess-ratefine-tuningauxiliary-module-trainingcomponent-scratch-training - GEAR-VLA: Learning Geometry-Aware Action Representations for Generalizable Robotic Manipulation
Qwen2.5-VL 기반 VLA에 latent action token K/V cache-conditioned stop-gradient DiT flow action expert, VGGT 기반 3D spatial encoder, relative end-effector action 기반 embodiment canonicalization을 결합해 unseen object / background shift / pretraining-unseen robot embodiment transfer를 개선하는 geometry-aware manipulation policy
Koreansuccess-rateVLAfine-tuningauxiliary-module-trainingcomponent-scratch-trainingcross-embodiment - MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation
Cosmos-Predict2.5 기반 Video DiT의 intermediate denoising feature를 Motion DiT action policy에 주입하고, SONIC 기반 unified whole-body motion token으로 humanoid의 상·하체를 한 action space에 묶어 Unitree G1에서 real-time loco-manipulation을 수행
Koreaninference-timesuccess-rateWAMfine-tuningcomponent-scratch-training - Q-VGM: Q-Guided Value-Gradient Matching for Flow-Matching VLA Policies
few-shot SFT된 π0.5 flow-matching VLA를 고정된 self-rollout buffer와 learned Q-critic의 action-gradient로 offline RL fine-tuning하되, Q-gradient를 terminal action label이 아니라 denoising-time residual velocity supervision으로 바꾸어 학습
KoreanVLAsuccess-ratefine-tuningauxiliary-module-training - ActionMap: Robot Policy Learning via Voxel Action Heatmap
VLA의 기존 single-point action decoder를 3D translation / 3D rotation / gripper voxel heatmap action head로 교체해, action space의 geometric proximity(인접성)를 학습 신호로 활용
Koreansuccess-rateVLAfine-tuningcomponent-scratch-training - Weekly Review #1
2026.05.19 ~ 2026.06.05
Korean - Flash-WAM: Modality-Aware Distillation for World Action Models
WAM의 video/action diffusion denoising을 각각의 noise regime에 맞게 다르게 distill해서, WAM을 거의 teacher 성능에 가깝게 유지하면서 real-time chunk-level control이 가능한 수준까지 가속하는 step-distillation method
KoreanWAMinference-timefine-tuningdistillation - 3DThinkVLA: Endowing Vision-Language-Action Models with Latent 3D Priors via 3D-Thinking-Guided Co-training
pretrained VLA를 VLA data + real-world 3D reasoning data로 co-training하면서, 3D foundation model과 reasoning-prompt teacher를 학습 중에만 사용해 2D image-only inference에서도 implicit 3D spatial reasoning을 action prediction에 주입
KoreanVLAsuccess-ratefine-tuningauxiliary-module-trainingcomponent-scratch-training - GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors
3D asset과 video foundation model prior를 이용해 humanoid loco-manipulation용 4D human-object interaction 데이터를 완전 디지털로 생성하고, 이를 Unitree G1용 tracking policy와 egocentric visual policy로 변환해 실제 로봇에 배포하는 data-generation / sim-to-real framework
Koreansuccess-ratetraining-datafine-tuningauxiliary-module-trainingsim2realcross-embodiment - OSCAR: Omni-Embodiment Skeleton-Conditioned World Action Model for Robotics
pretrained Cosmos-Predict2.5-2B video DiT를 2D kinematic skeleton condition으로 fine-tuning하여, 여러 robot embodiment와 human hand에 걸쳐 action-conditioned future video를 생성하고 이를 RoboArena policy evaluation proxy로 쓴다
KoreanWAMsuccess-ratefine-tuningcross-embodiment - Cosmos 3: Omnimodal World Models for Physical AI
language, image, video, audio, action을 하나의 Mixture-of-Transformers (MoT) 기반 omnimodal world model로 통합해, VLM·video generator·forward/inverse dynamics·robot policy를 하나의 Physical AI backbone으로 다루는 NVIDIA의 대규모 foundation model
KoreanWAMsuccess-ratefoundation-Model - Denoising Tells When to Replan: Denoising-Variance Adaptive Chunking for Flow-Based Robot Policies
last denoising step들에서 clean-action estimate들의 variance를 future action별 stability proxy로 사용해, 안정적인 action prefix만 실행하고 고분산 구간 전에 replan
KoreanVLAinference-timetraining-free - PointAction: 3D Points as Universal Action Representations for Robot Control
pretrained video diffusion model이 RGB뿐 아니라 temporally consistent XYZ pointmap까지 생성하게 만들고, 이 3D point dynamics를 embodiment-specific diffusion action decoder가 action chunk로 변환
KoreanWAMsuccess-ratefine-tuningcomponent-scratch-training - See Less, Specify More: Visual Evidence Budgets for Generalizable VLAs
VLA executor가 coarse goal과 full image에서 “무엇을 할지/무엇을 볼지”를 스스로 추론하지 않도록 goal-preserving local language와 learned visual evidence budget을 함께 학습시키는 planner-executor VLA generalization framework
KoreanVLAsuccess-ratefine-tuning - 3. Nsight Systems profiling & further optimization
Nsight Systems를 이용해서 bottleneck 지점을 더 정확하게 찾고 원인 분석 및 최적화
KoreanPythonProfilingNsight-Systems - Continuous Reasoning for Vision-Language-Action
VLA의 reasoning을 자연어 CoT가 아니라, 다른 VLA instance도 consume할 수 있는 WAE-regularized Gaussian continuous reasoning interface로 정의
KoreanVLAsuccess-ratefine-tuning - PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking
action chunking robot policy에서 고정 execution horizon 대신, predicted action chunk의 low-speed valley를 phase boundary로 사용해 매 query마다 실행 길이를 동적으로 선택하는 training-free test-time execution 방법
KoreanVLAinference-timetraining-free - VLAMotor: Test-Guided Enhancement of Vision-Language-Action Models via Agent-Based Data Synthesis
training distribution에서 멀고 서로 중복되지 않는 테스트 케이스로 VLA 실패를 적극적으로 찾고, 그 실패 trajectory를 VLM agent가 성공 trajectory로 고쳐 fine-tuning data로 쓰는 failure-driven VLA enhancement framework
KoreanVLAsuccess-ratefine-tuning - τ0-WM: A Unified Video-Action World Model for Robotic Manipulation
action generation, video prediction, action-conditioned evaluation을 하나의 shared video diffusion backbone 위에서 통합한 manipulation framework
KoreanWAMsuccess-ratefoundation-Model - Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments
Qwen3.5 VLM + DiT flow-matching action decoder / embodiment-aware prompt / joint pretraining → generalist VLA (manipulation, navigation, human egocentric motion, trajectory prediction)
KoreanVLAfoundation-Model - 0. VAE(Variational AutoEncoder)
DDPM의 variational perspective를 이해하는 데 필요한 VAE의 핵심 개념을 정리
KoreanWriting - ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models
VLA가 매 control step마다 전부 “생각”하지 않고, 현재 로봇 phase가 안정적인지/민감한지를 보고 Vision-LLM과 action head 계산을 동적으로 재사용하는 plug-in inference scheduler
KoreanVLAscheduler-training - SANTS: A State-Adaptive Scheduler for World Action Models
WAM이 매번 미래 영상을 끝까지 denoise하지 않고, 현재 로봇 상태에 따라 “여기서 멈출지”와 “얼마나 크게 건너뛸지”를 결정해 full-denoising WAM 대비 success-latency tradeoff를 개선하는 state-adaptive video denoising scheduler
KoreanWAMscheduler-training - A Factory-Floor Deployment Case Study of VLA Pipelines for Industrial Packaging Task: Workflow, Failures, and Lessons
데이터 수집·teleoperation·runtime·failure analysis 루프를 설계해서 pretrained π0.5를 실제 공장 포장 작업에 배포하는 시도, 그리고 거기서 얻은 교훈들
Koreanreal-worldVLAfine-tuning - HyperSim: A Holistic Sim-To-Real Framework For Robust Robotic Manipulation
더 현실적인 시뮬레이션 + 더 다양한 recovery trajectory + 소량 real data co-training → zero-shot/few-shot sim-to-real 성능 향상
Koreansim2realVLAfine-tuning - 2. Shallow-π Baseline Latency Check
Profiling tool들을 사용하기 전에 profiler 없는 순수 latency를 먼저 확인
KoreanPythonProfiling - SMoDP: Semantically Structured Mixture-of-Experts for Compositional Robotic Manipulation
Diffusion policy의 MoE router를 skill-aware하게 만들어 multi-task manipulation에서 expert를 의미 있는 skill 단위로 재사용하게 만든다
KoreanMoEmulti-taskdiffusion-policyscratch-training - OpenPI
ChatGPT랑 codex를 이용해서 openpi 레포지토리 분석하게 시켜보기
KoreanGraduate-School - Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs
π0-style flow-matching dVLA의 replanning latency를 lightweight draft와 flow-consistency verification으로 줄이는 speculative inference framework
KoreanVLAauxiliary-module-training - DEFLECT: Delay-Robust Execution via Flow-matching Likelihood-Estimated Counterfactual Tuning for VLA Policies
fresh observation에서 나온 action이 stale observation에서 나온 action보다 선호된다는 label-free preference pair를 이용해서 async VLA의 delay-robustness를 높이는 offline post-training 방법
KoreanVLAfine-tuningWriting - 1. Shallow-π implementation
π0 distillation을 통해 Shallow-π 구현 완료
KoreanPython - OxyGen: Unified KV Cache Management for VLA Inference under Multi-Task Parallelism
MoT VLA에서 action과 language task가 공유하는 observation KV cache를 통합 관리해 중복 prefill과 resource contention을 줄이고 action frequency와 language throughput을 동시에 높이는 inference system
KoreanVLAtraining-free