← Llull Lab· Paris research laboratory
As of 2 November 2026 · Updated three times a dayLlull Lab · Paris

AI & Robotics News

Research and policy news on artificial intelligence and robotics, updated three times a day from the public feeds of research laboratories, universities and EU institutions, with a daily selection of new research papers. Each item has a short summary and a link to the source.

This is the edition of 2 November 2026. See the latest news.

News

Updated three times a day
OpenAI2 November 2026

LifeVal

OpenAI announces LifeVal, a benchmark that the company describes as measuring how well AI systems help people get through ordinary daily tasks. The example tasks it gives include searching for childcare, handling personal finances and dealing with government paperwork.

Read the original at OpenAI ↗
OpenAI2 November 2026

Introducing the Grid Protectors Program

OpenAI announces the Grid Protectors Program, under which it says it will collaborate with electric utilities on future compute campuses. The stated aims are for these facilities to support the reliability of the power grid and to protect what consumers pay for electricity.

Read the original at OpenAI ↗

New research papers

arXiv

Preprints from the arXiv categories Robotics (cs.RO) and Artificial Intelligence (cs.AI). Preprints are not peer reviewed.

01
arXiv · cs.RO30 October 2026

VLA Policies Do Not Know When Not to Act: Unintended Task Execution and Start Suppression

The authors test six vision-language-action checkpoints in LIBERO-Goal with empty, nonsense, infeasible and wrong-task instructions, reporting that robots began moving in every empty-instruction episode and often finished unrequested tasks. They compare three suppression methods and report that an instruction prefix never stopped motion while an action-similarity gate held the robot in five checkpoints.

Read the original at arXiv ↗
02
arXiv · cs.RO30 October 2026

When the Simulator Fails, Not the Policy: Auditing Contact Artifacts in Simulated Robot Benchmarks

The authors present SimuGuard, an auditing plugin for several robot simulation benchmarks that flags objects ejected at contacts and replays episodes deterministically. They trace RoboTwin ejection artifacts to a single area threshold in collision detection and report that changing it removes the artifacts and raises success on five container tasks from 89% to 98%.

Read the original at arXiv ↗
03
arXiv · cs.AI30 October 2026

Subliminal Reinforcement Learning: traits travel through a judge's scores

The authors study whether a trait can pass to a student model trained with group relative policy optimization using only a biased judge's scores on semantically unrelated outputs such as invented syllables. They report detectable transfer of animal preferences and of undesirable safety-relevant dispositions, and describe a four-stage chain that they say anticipates where transfer occurs.

Read the original at arXiv ↗
04
arXiv · cs.RO30 October 2026

MolmoSpaces-CommonSense: Evaluating Commonsense in Robotic Manipulation Policies

The authors introduce a taxonomy of manipulation reasoning drawn from developmental cognitive science and a simulation benchmark testing occlusion, containment, support, affordances and hidden objects. Evaluating zero-shot generalist policies, they report success dropping by more than half as distractors grow and correct object-part grasping in only about half of cases.

Read the original at arXiv ↗
05
arXiv · cs.RO30 October 2026

ROOF: Learning Humanoid Fall Recovery under Overhead Constraints

The authors present a reinforcement learning method for humanoid fall recovery beneath overhead obstacles that rewards gradual reduction of robot-obstacle overlap, combined with a motion prior and support costs, using proprioception only. They report 88.5% simulated recovery under liftable plates and 95 successes in 150 real attempts on a Unitree G1.

Read the original at arXiv ↗
06
arXiv · cs.AI30 October 2026

Same Question, Different Answers: Auditing Persona-Dependent Variation in Six LLMs on Smoking-Cessation Advice

The authors audit six deployed LLMs on smoking-cessation questions across twelve user personas, reporting that teen personas were told about approved prescription medications far less often and that apparent experts often received fewer facts. They find a factual-floor system prompt narrowed persona gaps, though its effect on overall coverage differed across models.

Read the original at arXiv ↗
07
arXiv · cs.RO30 October 2026

Biomechanics do not predict perceptual and cognitive consequences of force augmentation

Using an elbow-assisting exosuit during lifting, the authors measure weight perception, adaptation to changing conditions and effort-based choices with and without assistance. They report that assistance lowered muscle activity, yet perceptual and decision effects depended on initial experience and did not simply follow the biomechanical reduction.

Read the original at arXiv ↗

AI-generated summaries. Items, titles and links come from the publishers' public news feeds and from arXiv, and each title links to the original. Summaries are written and checked automatically from the source text and are not individually reviewed by a person; the original is authoritative. Research papers are preprints, not peer reviewed. Llull Lab is independent of the publishers listed. How it is made.