Weekly Research Digest — 2026-07-06

17 new entries this week across 3 topic areas (includes July 2026 papers plus belated additions from April–June not previously logged).


Vision-Language-Action (VLA) Models

ReleaseVenueSignificance
neuro-symbolic-safety-guidance-vla-constrained-flow-matching Neuro-Symbolic Safety Guidance for VLA via Constrained Flow MatchingarXiv 2607.01378First trajectory-level safety framework for flow-matching VLAs; predictive collision avoidance
learning-action-priors-cross-embodiment-robot-manipulation Learning Action Priors for Cross-embodiment Robot ManipulationarXiv 2606.26095Two-stage motion prior pretraining before VLA alignment improves cross-embodiment generalization
last-hd-latent-physical-reasoning-scalable-human-data LaST-HD: Learning Latent Physical Reasoning from Scalable Human DataarXiv 2606.23685Aligns human-hand and robot demonstrations in shared latent space for scalable dexterous skill learning
apt-action-expert-pretraining-vla-instruction-generalization APT: Action Expert Pretraining Improves VLA Instruction GeneralizationarXiv 2606.12366Bayesian factorization + pretraining order fix for OOD language generalization in VLAs
vla-pro-cross-task-procedural-memory-transfer VLA-Pro: Cross-Task Procedural Memory TransferarXiv 2605.29562Runtime LoRA adapter retrieval from procedural memory bank for cross-task VLA generalization
intentvla-short-horizon-intent-modeling-aliased-robot-manipulation IntentVLA: Short-Horizon Intent Modeling for Aliased ManipulationarXiv 2605.14712Solves observation aliasing with history-conditioned intent representations + new AliasBench
loopvla-recurrent-refinement-vla-models LoopVLA: Recurrent Refinement for VLA ModelsarXiv 2605.0994845% smaller VLA model via looped Transformer; 1.7× faster inference

World Models for Robotics

ReleaseVenueSignificance
abot-m05-unified-mobility-manipulation-world-action-model ABot-M0.5: Unified Mobility-and-Manipulation WAMarXiv 2607.00678First WAM designed for combined navigation+arm control; addresses 3 structural mismatches in prior WAMs
roboworld-neural-simulators-generalist-robot-policy-evaluation RoboWorld: Neural Simulators for Robot Policy EvaluationarXiv 2607.01060World models as scalable policy evaluation tools without physical robot access
worldsample-closed-loop-real-robot-rl-world-modelling WorldSample: Closed-loop Real-robot RL with World ModellingarXiv 2607.02431Real-robot rollouts ground world model generation, reducing hallucination for RL augmentation
orca-world-foundation-model-multimodal-world-signals Orca: The World is in Your MindECCV 2026 / arXiv 2606.30534General world foundation model (BAAI) trained on 125K hours of video; unified Next-State-Prediction
himem-wam-hierarchical-memory-gated-world-action-models HiMem-WAM: Hierarchical Memory-Gated WAMarXiv 2606.10363Boundary-triggered memory gates at skill transitions enable robust long-horizon manipulation
echo-memory-controlled-study-memory-action-world-models Echo-Memory: Controlled Study of Memory in Action World ModelsarXiv 2606.09803First controlled ablation of memory design choices in world action models

Reinforcement Learning for Robotics

ReleaseVenueSignificance
one-demonstration-enough-real-world-robotic-rl One Demonstration Is Enough for Real-World Robotic RLarXiv 2607.01651AutoSERL: single demo matches 20-demo SERL baseline with automated interventions
aspire-agentic-skills-discovery-robotics ASPIRE: Agentic Skills Discovery for RoboticsarXiv 2607.00272NVIDIA-led system; 77% LIBERO-Pro gain, 31% zero-shot long-horizon vs. 4% prior
rho-coding-agent-roboticist-neurosymbolic-policies RHO: Your Coding Agent is Secretly a RoboticistarXiv 2606.16458Coding agents discover neurosymbolic robot policies via reward-based RL; new SOTA on Robosuite
last-r1-reinforcing-manipulation-adaptive-physical-latent-reasoning LaST-R1: Reinforcing Manipulation via Adaptive Physical Latent ReasoningarXiv 2604.28192LAPO jointly optimizes latent CoT reasoning + action generation; 99.9% LIBERO success rate

Generated automatically. All entries verified via web search.