Weekly Research Digest — 2026-06-22

14 new entries this week across 3 topic areas.


Vision-Language-Action (VLA) Models

ReleaseVenueSignificance
thinkingvla-interleaved-vision-language-reasoning ThinkingVLAarXiv 2026-06-16Forward+inverse CoT in a MoT architecture; strong gains on long-horizon tasks
finetuning-vla-fewer-layers Finetuning VLA Requires Fewer LayersarXiv 2026-06-18Training-free 50% layer pruning with no performance loss — pi0 and GR00T-N1.5 are over-parameterized
labvla-grounding-vla-scientific-laboratories LabVLAarXiv 2026-06-11First VLA for scientific lab automation; Qwen3-VL + DiT action expert
agentic-vla-efficient-online-adaptation Agentic-VLAarXiv 2026-05-21Agentic online adaptation with adaptive reward synthesis and experience memory
from-human-videos-to-robot-manipulation-survey From Human Videos to Robot ManipulationIJCAI 2026 SurveyTaxonomy of human-video→VLA learning: latent action, world models, 2D/3D supervision
dexora-open-source-vla-bimanual-dexterity DexoraICRA 2026First open-source dual-arm dual-hand VLA with 100K sim + 10K real episodes

World Models for Robotics

ReleaseVenueSignificance
memorywam-efficient-world-action-modeling-persistent-memory MemoryWAMarXiv 2026-06-18Hybrid persistent memory (recent+anchor+gist tokens) for non-Markovian WAMs; 70pp improvement
weaver-effective-world-model-robotic-manipulation WEAVERarXiv 2026-06-11Multi-view flow-matching world model; ρ=0.87 eval correlation, 38% policy improvement on pi0.5
world-models-robotic-manipulation-survey World Models for Robotic Manipulation: A SurveyarXiv 2026-06-01Three-axis taxonomy: representation predicted, action coupling, pipeline stage
world-action-verifier-self-improving-forward-inverse-asymmetry World Action Verifier (WAV)ICLR 2026 Outstanding PaperSelf-improving world models via forward-inverse asymmetry; no ground truth labels needed
playworld-robot-world-models-autonomous-play PlayWorldarXiv 2026-03-09World model trained from autonomous robot self-play; 65% policy improvement vs. human demos

Reinforcement Learning for Robotics

ReleaseVenueSignificance
playful-agentic-robot-learning-rats Playful Agentic Robot Learning (RATs)arXiv 2026-06-17Self-directed play builds reusable code skill library; +20pp on LIBERO-PRO vs. no-play baseline
rove-human-interventions-humanoid-manipulation-rl ROVEarXiv 2026-06-15Optimistic Value Estimation for RL from imperfect human interventions on humanoid VLAs
mpc-guided-rl-humanoid-locomotion-manipulation MPC-Guided RL for HumanoidarXiv 2026-06-04GPU-native parallel MPC reward signal makes MPC-guided RL practical at scale for humanoids

Generated automatically. All entries verified via web search.