Weekly Research Digest — 2026-08-23

29 new entries this week across 4 topic areas.


Vision-Language-Action (VLA) Models

ReleaseVenueSignificance
eximo-vlm-guided-exploration-vla-policies EXIMO: VLM Guided Exploration of VLA PoliciesarXiv⭐ HIGH PRIORITY: VLM-planned autonomous exploration replaces teleoperated data collection for VLA post-training
tau0-vla-hierarchical-robot-foundation-model-test-time-computation τ0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time ComputationarXiv⭐ HIGH PRIORITY: confidence-gated world-model search allocates extra test-time compute only where needed
stellavla-in-context-structured-demonstration-generalizable-vla StellaVLA: In-Context Structured Demonstration for Generalizable VLAarXiv⭐ HIGH PRIORITY: annotation-free in-context conditioning tops the VLA-Arena leaderboard, no fine-tuning needed
imagining-recovery-inference-time-counterfactual-realignment-vla Imagining Recovery: Inference-Time Counterfactual Realignment for VLAarXiv⭐ HIGH PRIORITY: training-free recovery from goal/scene disruptions via imagined counterfactual continuation
reflex-fast-predictive-vla-reaction-critical-manipulation Reflex: Enabling Fast and Predictive VLA Models for Reaction-Critical ManipulationarXivNew latency-aware benchmark (ReflexBench) + efficient policy for dynamic manipulation, 65ms deployment latency
foca-future-oriented-conditioning-data-efficient-vla-adaptation FOCA: Future-Oriented Conditioning for Data-Efficient VLA AdaptationarXiv / ICML 2026⭐ HIGH PRIORITY: new SOTA few-shot adaptation, 95.7% success on LIBERO with only 20 demos
bridging-morphology-gap-intent-conditioned-fine-tuning-dexterous-vla Bridging the Morphology Gap: Intent-Conditioned Fine-Tuning for Dexterous VLAarXiv⭐ HIGH PRIORITY: data-efficient fine-tuning recipe for moving low-DoF-pretrained VLAs onto dexterous hands
gen-15-embodied-foundation-models-one-shot-learners GEN-1.5: Embodied Foundation Models are One-Shot Learnersblog⭐ HIGH PRIORITY: Generalist AI claims emergent one-shot task learning from a single demonstration, no fine-tuning (unverified beyond blog claims)
hidden-in-plain-sight-diffusion-based-unrestricted-attacks-vla Hidden in Plain Sight: Diffusion-Based Unrestricted Robotic Attacks on VLA ModelsarXivAdversarial-robustness paper using diffusion-generated (rather than norm-bounded) perturbations against VLA policies

World Models for Robotics

ReleaseVenueSignificance
egowam-world-action-models-beyond-pixels-egocentric-human-data EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human DataarXiv⭐ HIGH PRIORITY: controlled study finds DINO/3D-flow targets beat raw pixel prediction by up to 4x OOD for human-video-to-robot transfer
r2rdreamer-3d-aware-data-augmentation-spatially-generalized-manipulation R2RDreamer: 3D-aware Data Augmentation for Spatially-generalized 2D Manipulation PoliciesarXiv⭐ HIGH PRIORITY: real-to-real 3D demonstration editing + video completion for spatial generalization
halo-wa-hybrid-attention-latent-guided-online-rl-world-action-models HALO-WA: Hybrid-Attention Latent-Guided Online RL for World-Action ModelsarXiv⭐ HIGH PRIORITY: online RL adapter reading WAM internals raises precision-manipulation success from 26.4% to 87%
wh0-generative-world-models-scalable-egocentric-hand-manipulation-data Wh0: Generative World Models as Scalable Sources of Egocentric Human Hand Manipulation DataarXiv⭐ HIGH PRIORITY: world model as synthetic-data engine, zero-shot dexterous success 8.3%→38.9% across 18 real tasks
wla-0-world-language-action-model-unified-world-modeling World-Language-Action Model (WLA-0)arXivAutoregressive-Transformer alternative to the now-dominant diffusion-Transformer WAM architecture
paiworld-3d-consistent-world-foundation-model-manipulation PAIWorld: A 3D-Consistent World Foundation Model for Robotic ManipulationarXivExplicit multi-view geometric reasoning fixes cross-view drift/depth inconsistency in multi-camera WAMs
physisforcing-physics-reinforced-world-simulator-manipulation PhysisForcing: Physics Reinforced World Simulator for Robotic ManipulationarXivPixel + semantic alignment losses raise closed-loop world-model success 16.0%→24.0%
dim-wam-world-action-modeling-diverse-historical-event-memory DIM-WAM: World-Action Modeling with Diverse Historical Event MemoryarXivMulti-scale memory banks raise RMBench average success from 28.4% to 69.8% on long-horizon tasks

Reinforcement Learning for Robotics

ReleaseVenueSignificance
temporal-grpo-beyond-trajectory-level-credit-vla-rl Temporal GRPO: Beyond Trajectory-Level Credit in VLA RLarXiv⭐ HIGH PRIORITY: step-level credit assignment for GRPO-based VLA RL fine-tuning
recovla-vlm-guided-reward-compilation-failure-recovery-vla ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in VLA PoliciesarXiv⭐ HIGH PRIORITY: frozen-base + residual RL recovery, validated zero-shot on real Fetch hardware
score-support-constrained-rl-real-world-policy-improvement SCORE: Support-Constrained RL Enables Real-World Policy Improvement without Real-World ExperiencearXiv⭐ HIGH PRIORITY: flow-steered support-constrained sim RL improves real dexterous manipulation with zero extra real data
robot-self-improvement-human-video-dynamics-models Robot Self-Improvement via Human-Video Dynamics ModelsarXiv⭐ HIGH PRIORITY: training-free failure correction via human-video dynamics model, 40%→81% across 7 tasks and multiple backbones
regrind-minimalist-retargeting-guided-rl-dexterous-manipulation REGRIND: A Minimalist Retargeting-Guided RL Recipe for Dexterous ManipulationarXivObject-centric keypoint retargeting + residual RL for contact-rich tool-use, with sim-to-real transfer analysis

Humanoid Robotics

ReleaseVenueSignificance
toward-certified-functional-safety-industrial-humanoid-robots Toward Certified Functional Safety for Industrial Humanoid RobotsarXivIdentifies the “fail-passive gap”: legged humanoids violate standard industrial safety-certification assumptions
whole-body-planning-humanoids-confined-spaces-self-collision-avoidance Whole-Body Planning for Humanoids Navigating Confined SpacesarXivThree-stage plan-then-residual-RL pipeline for humanoid navigation through tight, cluttered spaces
humantracker-comprehensive-human-aligned-motion-tracking-benchmark HumanTracker: Comprehensive Human-Aligned Motion Tracking BenchmarkarXiv~153-hour benchmark + preference-aligned HumanScore metric catches artifacts kinematic-error metrics miss
noitom-hiphi-617-hour-high-precision-human-motion-dataset Noitom Robotics Releases HiPHI: 617-Hour High-Precision Human Motion DatasetblogLarge public studio-precision mocap + interaction dataset for humanoid imitation learning
adapt-humanoid-robots-professional-style-tennis AdaPT: Humanoid Robots Learn Professional-Style TennisblogStyle-specific motion imitation from broadcast footage, in-the-wild serving on Unitree G1 / Dobot Atom
unitree-superman-humanoid-record-jump-sprint-speed Unitree Unveils “Superman” Humanoid with Record Jump and Sprint SpeedblogUnverified hardware showcase (2m jump, 12.66 m/s sprint) timed with Unitree’s Shanghai IPO
agility-robotics-digit-v5 Agility Robotics Unveils Digit V5blogFence-free industrial humanoid targeting Dec. 2026 deployment, alongside Agility’s $2.5B SPAC listing

Generated automatically. All entries verified via web search.