← Llull Lab· Paris research laboratory
As of 31 October 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 31 October 2026. See the latest news.

News

Updated three times a day
OpenAI30 October 2026

Fair use benefits everyone

OpenAI highlights portions of court briefs submitted by creators, scholars, accessibility advocates, and legal professionals that argue fair use doctrine supports the training of AI systems. The shared excerpts collectively make the case that fair use in this context promotes innovation and provides broad societal benefits.

Read the original at OpenAI ↗
MIT News30 October 2026

MIT affiliates named 2026 Schmidt Sciences AI2050 Fellows

Schmidt Sciences named five MIT faculty members and two alumni as 2026 AI2050 Fellows. Among them, Senior Fellow Sendhil Mullainathan, who holds dual appointments in Economics and Electrical Engineering and Computer Science at MIT, is described as focusing on human-centered AI and launching The Bike Shop@MIT, a research center working on scalable cognitive tools and new machine learning approaches.

Read the original at MIT News ↗
ETH Zurich29 October 2026

What’s that crawling around?

ETH Zurich researchers present a robotic hand capable of crawling autonomously over surfaces such as a piano and through corridors. The project, demonstrated in a Halloween-themed video, showcases the dexterous locomotion capabilities of the hand-shaped robot and the research questions driving its development.

Read the original at ETH Zurich ↗
European Commission29 October 2026

Local Digital Twins CiTiVERSE EDIC joins Smart City Expo World Congress 2026

The European Commission describes the Local Digital Twin CiTiVERSE European Digital Infrastructure Consortium, which brings together city, regional, and industrial stakeholders to develop AI solutions for urban governance and public services. The initiative, aligned with the EU Apply AI Strategy, aims to build an ecosystem of human-centric AI and digital twin tools that cities can replicate.

Read the original at European Commission ↗
MIT News30 October 2026

Q&A: Hey AI, who should I vote for this year?

MIT researchers have created the LLM Election Observatory, a real-time dashboard that tracks how popular AI chatbots respond to questions about the 2026 U.S. midterm elections. Led by Professors Berinsky, Podimata, and Stewart, the team periodically reruns a fixed set of politically themed questions to monitor how chatbot answers shift based on factors like user location and party affiliation.

Read the original at MIT News ↗

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.AI29 October 2026

Efficient Scalable Oversight via Rewinding

The authors introduce rewindable interactive proofs, a new model for scalable oversight where a weak verifier can save and restore an AI prover's state during interaction. They construct a protocol for oracle-aided computations with robust outputs where the honest prover runs in linear time and the verifier runs in near-linear time in input length, using no cryptography, and they report an implementation.

Read the original at arXiv ↗
02
arXiv · cs.RO29 October 2026

Cross-Controller Transfer via In-Context Behavior Adaptation

The authors study transferring robot data across embodiments and controllers by encoding behavioral interactions through examples of commanded actions and resulting motions, using a transformer context encoder for in-context adaptation. They report 33 percentage point improvement in average success over naive data pooling, with 50% success on unseen controller modes and effective adaptation on contact-rich tasks like USB insertion.

Read the original at arXiv ↗
03
arXiv · cs.RO29 October 2026

Continuous-Action World Models: Learning Cross-Embodiment Spatiotemporal Dynamics Priors

The authors propose Continuous-Action World Models (CoAWM), a framework learning cross-embodiment spatiotemporal dynamics priors through a shared latent function space over robot actions. They describe a two-stage approach where function pretraining aligns heterogeneous robot trajectories into a common representation, and a world-model backbone then predicts latent action functions conditioned on observations and language instructions.

Read the original at arXiv ↗
04
arXiv · cs.RO29 October 2026

Humanoid Dodgeball: Learning Fast Dynamic Obstacle Avoidance Motions for Humanoids

The authors present Humanoid Dodgeball, a reinforcement learning framework for agile dynamic obstacle avoidance in humanoid robots using threat-aware reward shaping and direction-specific adversarial motion priors. They report that the framework substantially improves evasive performance against incoming projectiles across diverse speeds and distances in both simulation and hardware experiments while preserving physically plausible whole-body motions.

Read the original at arXiv ↗
05
arXiv · cs.AI29 October 2026

Proving and Improving the Platonic Representation Hypothesis with JEPAs

The authors prove a Platonic Representation Theorem showing that Joint-Embedding Predictive Architectures trained separately on different modalities learn the same representation up to rotation when positive pairs are observations of one scene. They report that a 109M-parameter text JEPA aligns with vision encoders better than every evaluated language model up to 72B parameters, suggesting next-token prediction is an inefficient route to platonic representations.

Read the original at arXiv ↗
06
arXiv · cs.RO29 October 2026

SwapBot: Tool-Centric Robot Manipulation via Interchangeable End-Effectors and Tactile Sensing

SwapBot is a modular end effector system that pairs a fixed robot-side coupling with inexpensive tool-side adapters, turning common household tools into interchangeable robot end effectors equipped with compact tactile sensors. User studies found it lowers teleoperation time and workload, and imitation learning policies trained with tactile-augmented demonstrations achieved 15 to 25 percentage point higher success rates than RGB-only policies on contact-rich cooking tasks.

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.