Weekly Research Digest — 2026-06-22
14 new entries this week across 3 topic areas.
Vision-Language-Action (VLA) Models
| Release | Venue | Significance |
|---|---|---|
| thinkingvla-interleaved-vision-language-reasoning ThinkingVLA | arXiv 2026-06-16 | Forward+inverse CoT in a MoT architecture; strong gains on long-horizon tasks |
| finetuning-vla-fewer-layers Finetuning VLA Requires Fewer Layers | arXiv 2026-06-18 | Training-free 50% layer pruning with no performance loss — pi0 and GR00T-N1.5 are over-parameterized |
| labvla-grounding-vla-scientific-laboratories LabVLA | arXiv 2026-06-11 | First VLA for scientific lab automation; Qwen3-VL + DiT action expert |
| agentic-vla-efficient-online-adaptation Agentic-VLA | arXiv 2026-05-21 | Agentic online adaptation with adaptive reward synthesis and experience memory |
| from-human-videos-to-robot-manipulation-survey From Human Videos to Robot Manipulation | IJCAI 2026 Survey | Taxonomy of human-video→VLA learning: latent action, world models, 2D/3D supervision |
| dexora-open-source-vla-bimanual-dexterity Dexora | ICRA 2026 | First open-source dual-arm dual-hand VLA with 100K sim + 10K real episodes |
World Models for Robotics
| Release | Venue | Significance |
|---|---|---|
| memorywam-efficient-world-action-modeling-persistent-memory MemoryWAM | arXiv 2026-06-18 | Hybrid persistent memory (recent+anchor+gist tokens) for non-Markovian WAMs; 70pp improvement |
| weaver-effective-world-model-robotic-manipulation WEAVER | arXiv 2026-06-11 | Multi-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 Survey | arXiv 2026-06-01 | Three-axis taxonomy: representation predicted, action coupling, pipeline stage |
| world-action-verifier-self-improving-forward-inverse-asymmetry World Action Verifier (WAV) | ICLR 2026 Outstanding Paper | Self-improving world models via forward-inverse asymmetry; no ground truth labels needed |
| playworld-robot-world-models-autonomous-play PlayWorld | arXiv 2026-03-09 | World model trained from autonomous robot self-play; 65% policy improvement vs. human demos |
Reinforcement Learning for Robotics
| Release | Venue | Significance |
|---|---|---|
| playful-agentic-robot-learning-rats Playful Agentic Robot Learning (RATs) | arXiv 2026-06-17 | Self-directed play builds reusable code skill library; +20pp on LIBERO-PRO vs. no-play baseline |
| rove-human-interventions-humanoid-manipulation-rl ROVE | arXiv 2026-06-15 | Optimistic Value Estimation for RL from imperfect human interventions on humanoid VLAs |
| mpc-guided-rl-humanoid-locomotion-manipulation MPC-Guided RL for Humanoid | arXiv 2026-06-04 | GPU-native parallel MPC reward signal makes MPC-guided RL practical at scale for humanoids |
Generated automatically. All entries verified via web search.