Weekly Research Digest — 2026-07-13
12 new entries this week across 3 topic areas (July 7–13 papers plus belated additions from June not previously logged).
Vision-Language-Action (VLA) Models
| Release | Venue | Significance |
|---|---|---|
| lift3d-vla-3d-geometry-dynamics-aware-vla Lift3D-VLA | arXiv 2607.06564 | 3D point-cloud VLA with GC-MAE dual-objective SSL; +10.8%/+11.1% on MetaWorld/RLBench over best prior VLA |
| lingbot-vla-20-open-source-cross-embodiment LingBot-VLA 2.0 (Robbyant/Ant Group) | blog release | 6B cross-embodiment VLA on 60K hrs / 20 morphologies / 17 manufacturers; <130ms inference, open-source |
| lerobot-v060-world-model-policies-open-robotics LeRobot v0.6.0 (Hugging Face) | blog / toolkit release | VLA-JEPA + FastWAM + LingBot-VA world model policies; closes the evaluate-correct-train loop in open robotics |
| policytrim-intrinsic-policy-efficiency-vla-rl PolicyTrim | arXiv 2606.22540 | RL post-training reduces VLA inference frequency without architectural changes or new demos |
| embodied-r15-evolving-physical-intelligence-foundation-models Embodied-R1.5 | arXiv 2606.11324 | 8B EFM on Qwen3-VL; SOTA on 16/24 embodied benchmarks, fine-tunes to VLA beating π₀.₅ |
| embodied-cpp-portable-inference-runtime-vla-wam Embodied.cpp | arXiv 2607.02501 | Portable C++ runtime for VLA/WAM deployment on CPU/GPU/NPU; latency-first closed-loop design |
World Models for Robotics
| Release | Venue | Significance |
|---|---|---|
| rynnworld-4d-4d-embodied-world-model-robotic-manipulation RynnWorld-4D (Alibaba) | arXiv 2607.06559 | First world model co-generating RGB+depth+optical flow in one denoising loop; strong bimanual Sim2Real |
| rynnworld-teleop-action-conditioned-world-model-digital-teleoperation RynnWorld-Teleop (Alibaba) | arXiv 2607.06558 | Digital teleoperation: replaces physical robot with generative world model for data synthesis at 40+ FPS |
| dswam-dual-system-world-action-foundation-model-fine-grained-manipulation DSWAM | arXiv 2607.04927 | Dual-system WAM: System 1 executor + optional System 2 VLM planner for complex instruction decomposition |
| from-world-models-to-world-action-models-tutorial-robotics WM→WAM Tutorial | arXiv 2607.00836 | Concise design-space taxonomy covering 4 WAM paradigms; clearest conceptual roadmap for the WM→WAM transition |
Reinforcement Learning for Robotics
| Release | Venue | Significance |
|---|---|---|
| dvla-rl-reinforcement-learning-denoising-trajectories-discrete-diffusion-vla dVLA-RL | arXiv 2606.23623 | PPO-style RL for discrete diffusion VLAs via denoising-trajectory MDP; SOTA on LIBERO |
| taccorl-tactile-feedback-vla-simulation-cotraining TacCoRL | arXiv 2606.11743 | Tactile feedback in VLAs via sim-real co-training + RL on contact rollouts; no large-scale tactile pretraining needed |
Generated automatically. All entries verified via web search.