This is the edition of 8 November 2026. See the latest news.
News
Updated three times a day
Hugging Face·6 November 2026
A Hugging Face post presents ToyBlocks, which it describes as a controlled world built for training robots and evaluating how they learn. The candidate entry carries only this title and no body text, so the environment's design, tasks, benchmarks and any reported results cannot be described from the material supplied.
Read the original at Hugging Face ↗
Carnegie Mellon Robotics Institute·6 November 2026
Carnegie Mellon's Team Chiron, working with Lockheed Martin and the University of Pittsburgh, is building autonomous systems in which aerial drones spot injured people and pass reports to ground robots that collect more details. Robotics scientist Kimberly Elenberg said the aim is to help medics decide quickly whom to treat first during mass casualty events, as the team competes in a three-year DARPA contest.
Read the original at Carnegie Mellon Robotics Institute ↗
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 Preprints are not peer reviewed.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.AI5 November 2026
The authors propose SwarmEvolve, an evolutionary program search framework in which many inexpensive agents reason in latent space, exchange KV caches, and are filtered for novelty using hidden state distances. On 21 mathematical and scientific computing problems, they report the highest average score and faster convergence than competing frameworks.
Read the original at arXiv ↗
02
arXiv · cs.AI5 November 2026
The authors present SAVeR2, a safety alignment method for large reasoning models that splits safety into four verifiable reward signals and uses gradient projection to reduce conflicts between safety and reasoning objectives. Across models from 1.5B to 671B parameters, they report better safety and less over-refusal without loss of general reasoning.
Read the original at arXiv ↗
03
arXiv · cs.RO5 November 2026
The authors propose a kinematic hypothesis context, a signal both sender and receiver build from joint states, robot model, and calibration, which conditions a learned video codec for remote robot policy uplinks. On LIBERO they report bit rate savings over standard codecs and higher closed-loop policy success at low bandwidth.
Read the original at arXiv ↗
04
arXiv · cs.AI4 November 2026
The authors introduce Behavior-Conditioned Symbolic Synthesis, an automated search for interpretable deterministic programs that reproduce observed behavior, and apply it to human and machine responses on hundreds of reasoning problems. They report that both humans and models oversimplify and adopt partial rules, which they relate to satisficing.
Read the original at arXiv ↗
05
arXiv · cs.AI5 November 2026
The authors study whether benchmark gains from agent harness evolution reflect reliable, efficient, transferable improvement, evaluating four methods with repeated runs and interventions. They report that every evolved agent loses some previously solved tasks, that reliability gains often raise cost, and they propose EvoAudit to revise the evolution procedure.
Read the original at arXiv ↗
06
arXiv · cs.RO5 November 2026
The authors present TactileLab, an IsaacLab-based tactile simulation framework that estimates shear-related contact motion as simulated tactile flow using rigid-body physics instead of soft-body models. They also propose a real-to-sim translation method mapping real sensor images to simulated flow and depth, and report sim-to-real dexterous manipulation.
Read the original at arXiv ↗
07
arXiv · cs.AI5 November 2026
The authors analyze a sequential social learning model in which agents may use costly accurate manual fact-checking or a free but error-prone automated tool. They prove that the tool lowers the probability of correct cascades for articles of intermediate reliability under certain cost and informativeness conditions.
Read the original at arXiv ↗
08
arXiv · cs.AI4 November 2026
The authors build a theoretical framework contrasting Darwinian, Lamarckian, and agent-directed evolution after environmental shifts, finding that with effective learning each yields faster adaptation and that ideal search effort hinges on learning ability, difficulty, and experiment costs. They caution that a misspecified evaluation target lets extra search speed convergence toward wrong behavior, arguing that independent oversight must match autonomous AI progress.
Read the original at arXiv ↗
09
arXiv · cs.RO5 November 2026
MessyNav is a zero-shot framework using vision-language models that reasons about object affordances, choosing to swipe, lift, or steer around obstacles, before producing a high-level plan for interactive navigation in cluttered spaces. Tested on a mobile manipulator indoors and outdoors, it reportedly beat the best baseline by 27 percentage points in real trials without any task-specific training.
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.