A new framework called RoboFAC targets a key gap in robotic manipulation: Vision-Language-Action models are trained mainly on successful demos, leaving them brittle when things go wrong. The authors back it with 9,440 failure…

#RoboFAC #Robotics #VLA #AIResearch
https://arxiv.org/abs/2505.12224

A new arXiv paper shows that for vision-language-action models like π0.5 and GR00T N1.5, applying image augmentation only to the critic during reinforcement learning post-training boosts out-of-distribution success…

#Robotics #VLA #ReinforcementLearning #Robustness
https://arxiv.org/abs/2610.05994

New research defines correctability and actionability as evaluation axes for safety interventions on chain-of-thought in VLA policies, and introduces TRUST, a value model that scores partial reasoning prefixes to monitor and steer…

#AI #Robotics #VLA #AIsafety
https://arxiv.org/abs/2610.00601

Recursive Harness Distillation lets a strong agent distill intervention experience into a playbook that a lighter agent can reuse across new robot manipulation tasks, lifting real-world success from 37.3% to 64.0%.

#Robotics #VLA #Manipulation #AI
https://arxiv.org/abs/2609.33378

New work on arXiv shows that even brief, one-step observation perturbations can substantially degrade vision-language-action model performance, and proposes CARE to adapt execution length based on action chunk consistency. This…

#AI #Robotics #Safety #VLA
https://arxiv.org/abs/2609.32550

Shots of the NRAO Very Large Array near Socorro, New Mexico. These were taken in September, 2014 when the radio telescopes were in a close configuration.

#photo #photos #photography #astronomy #space #astrophotography #spacephotography #radiotelescopes #vla #nrao #nm #newmexico

OpenAI GPT-6 Astra: Semantic-to-Physical Manipulation and the Emerging Robotics Attack Surface

https://openai.robocurve.org/gpt-6-astra
https://flagthis.com/tldr/7033

##OpenAI ##EmbodiedAI ##RoboticsSecurity ##PromptInjection ##VLA